Paper deep dive
Habermolt: Delegating Deliberation to AI Representatives
Joseph Low, Oscar Duys, Claude Formanek, Michiel Bakker, Lewis Hammond
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%
Last extracted: 7/8/2026, 7:41:55 PM
Summary
The paper introduces AI-delegated deliberation, a paradigm where AI agents represent human users in collective decision-making to overcome human bandwidth constraints. The authors deploy and empirically evaluate HABERMOLT, a public platform implementing this paradigm across three dimensions: representation, aggregation, and revision. Findings indicate that autonomous AI participation yields less diverse opinions than human-directed interviews, and longer user memory profiles do not guarantee more distinctive outputs. The platform utilizes the Schulze ranking method for aggregation and a lazy consensus mechanism for asynchronous revision, highlighting key design challenges for scalable, trustworthy AI representatives in democratic contexts.
Entities (8)
Relation Signals (7)
HABERMOLT â implements â AI-delegated deliberation
confidence 95% ¡ To study these dynamics empirically, we deploy Habermolt, a public platform for AI-delegated deliberation.
HABERMOLT â uses â Schulze method
confidence 94% ¡ a winner is determined via the Schulze method over the resulting ranking distribution
HABERMOLT â employs â Lazy consensus
confidence 93% ¡ We refer to this as lazy consensus: the collective output reflects the current state of all agents at all times
HABERMOLT â supports â Autonomous participation
confidence 92% ¡ Autonomous participation is what gives HABERMOLT its scaling properties
OpenClaw â integrateswith â HABERMOLT
confidence 88% ¡ Users can either create an agent on the HABERMOLT platform or run their own Open-Claw agent; the platform interacts with both agent types identically through a shared API.
Habermas machine â comparesto â HABERMOLT
confidence 85% ¡ This contrasts with the Habermas Machine, where contributions are fixed once submitted within a round
Pol.is â comparesto â HABERMOLT
confidence 85% ¡ Unlike Pol.is which surfaces a descriptive picture of disagreement, the aim is a shared position the group can act on.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Deliberative democracy arguably leads to better collective decisions, but is fundamentally constrained by human attention and bandwidth. While recent AI-mediated deliberations scale participation by synthesizing inputs from many humans, they remain time-intensive for individual users. As AI models become increasingly capable, AI systems are being deployed not only to mediate deliberation between humans, but to represent humans in it: where AI agents deliberate on behalf of human users. We call this paradigm AI-delegated deliberation. While it promises unprecedented scale for democratic participation, it introduces qualitatively new design and alignment challenges that are poorly understood and under-theorized. To study these dynamics empirically, we deploy Habermolt, a public platform for AI-delegated deliberation. We evaluate its effectiveness along three dimensions that we use to organize any deliberative system: representation, aggregation, and revision. We use these observations to illuminate the design decisions future AI-delegated deliberation platforms must confront, contributing to the broader research agenda for scalable yet trustworthy AI representatives.
Tags
Links
- Source: https://arxiv.org/abs/2605.24413v3
- Canonical: https://arxiv.org/abs/2605.24413v3
Trouble viewing inline? Open PDF directly â
Full Text
60,638 characters extracted from source content.
Expand or collapse full text
HABERMOLT: Delegating Deliberation to AI Representatives Joseph Low 1 Oscar Duys 1 Claude Formanek 2 Michiel Bakker 3 Lewis Hammond 4 Abstract Deliberative democracy arguably leads to bet- ter collective decisions, but is fundamentally constrained by human attention and bandwidth. While recent AI-mediated deliberations scale par- ticipation by synthesizing inputs from many hu- mans, they remain time-intensive for individual users. As AI models become increasingly capa- ble, AI systems are being deployed not only to mediate deliberation between humans, but to rep- resent humans in it: where AI agents deliberate on behalf of human users. We call this paradigm 1 Cooperative AI Research Fellowship 2 AI Safety South Africa 3 MIT 4 Cooperative AI Foundation.Correspondence to: Joseph Low<jolow999@gmail.com>, Oscar Duys<oscar- duys@gmail.com>. Preprint. May 28, 2026. AI-delegated deliberation. While it promises un- precedented scale for democratic participation, it introduces qualitatively new design and alignment challenges that are poorly understood and under- theorized. To study these dynamics empirically, we deploy HABERMOLT, a public platform for AI-delegated deliberation. We evaluate its effec- tiveness along three dimensions that we use to organize any deliberative system: representation, aggregation, and revision. We use these observa- tions to illuminate the design decisions future AI- delegated deliberation platforms must confront, contributing to the broader research agenda for scalable yet trustworthy AI representatives. 1 arXiv:2605.24413v3 [cs.CY] 27 May 2026 HABERMOLT: Delegating Deliberation to AI Representatives 1. Introduction How do groups of people make decisions together? At a small scale, direct participation works: everyone speaks, everyone votes. But as groups grow, this breaks down. The solution that humans have converged on across centuries of political experimentation is representation: rather than par- ticipating directly, you delegate to someone who participates on your behalf. Representation solves the scale problem but creates a le- gitimacy gap: decisions are made for people rather than by them. Deliberative democracy (Habermas, 1996) par- tially closes this gap through forums like citizen assemblies, where a representative sample of selected citizens reason through contentious issues and produces recommendations often regarded as more legitimate than those of ordinary leg- islatures (Fishkin, 2009; OECD, 2020). Yet these fora face hard limits. They are expensive and slow to convene, and can engage only a small slice of the population at any one time. Scaling deliberative democracy, while preserving its epistemic and normative virtues, remains an open problem. Recent advances in large language models have opened a new path: using AI to scale deliberation beyond the constraints of human bandwidth. The most developed ap- proaches use AI as a mediator between participants who remain actively present, but this still leaves the population that can meaningfully engage bounded by human atten- tion. A more ambitious role for AI is that of representative: a persistent agent, initialized from a humanâs views, that participates on their behalf when the human cannot. This extends a userâs reach to decisions they would otherwise have had no say in, but introduces failure modes that are not yet well understood. To better understand the dynamics of this phenomenon we call AI-delegated deliberation, we publicly deployed a plat- form where AI agents deliberate on behalf of humans. In this technical report, we distill our experience from this and make two contributions: â˘HABERMOLT, a deployed platform for AI-delegated deliberation (Section 3). Each userâs agent learns their views through an interview, participates in delib- erations when the user is absent, which produces an output the user can inspect and correct at any time. ⢠A case analysis of HABERMOLT (Section 4). We ex- amine HABERMOLT empirically along the three dimen- sions of representation, aggregation and revision, eval- uating whether its design choices hold up in practice and where they do not. We then use these analyses to il- luminate the design landscape that future AI-delegated deliberation platforms must navigate. 2. Related Work Prior work on using AI to support collective deliberation has developed along two largely separate tracks, summarised in Figure 1. The first treats AI as a mediator that helps humans deliberate more effectively, though the human side of participation remains just as time-intensive. The second treats AI as a generative simulacrum that simulates human interaction with no human in the loop, typically to test hypotheses or explore counterfactuals rather than to reach a decision-relevant consensus. Each track addresses a real limitation of human-only deliberation, but neither closes the loop that AI-delegated deliberation requires: a human authoring their own input, an agent acting on their behalf when they are not present, and the human retaining the ability to inspect and correct what was said in their name. We review the two tracks in turn before returning to the gap between them. Figure 1. Prior work has developed along two tracks that do not yet meet. AI-delegated deliberation occupies the intersection: it inherits the scaling of simulation and the human-correction loop of mediated deliberation, but introduces its own distinctive limitations we outline in Section 4. 2.1. AI-Mediated Deliberation A growing body of work explores using AI to help groups of humans reach collective decisions more effectively. Tessler et al. (2024) introduce the Habermas Machine, an AI me- diator that generates group consensus statements and out- performs human mediators in preference ratings across a large-scale UK study. Tessler et al. (2026) extend this work by framing AI-mediated deliberation as navigating Fishkinâs trilemma among participation, quality, and equality. Pol.is (Small et al., 2021) takes a complementary approach, using pairwise voting and clustering to surface the geometry of disagreement rather than distilling a single consensus state- ment. Fish et al. (2025) provide a theoretical foundation for this cluster, introducing representation guarantees for selecting slates of statements that proportionally represent free-form opinions. These systems differ in how they structure participation and aggregate output, but they share a genuine achievement: 2 HABERMOLT: Delegating Deliberation to AI Representatives AI mediators can synthesise across hundreds or thousands of opinions at a scale no human facilitator could match, surfacing structure in disagreement and distilling consensus far more efficiently. The limitation is not on the machine side but on the human side. Participants must still read, write, and return as the discussion evolves, so engagement remains bounded by human attention. The AI accelerates the mediator but it does not reduce the cost for participants. 2.2. Generative Simulacra A second line of work asks what happens when an AI system stands in for a human entirely. The seminal example is Park et al. (2023), who instantiate generative agents in a Sims- like sandbox town and observe emergent social behaviour without any human involvement. More recently, MOLT- BOOK (Moltbook, 2026) has taken this further into the wild with a Reddit-style platform populated by autonomous AI agents that post, comment, and form communities, with humans relegated to observers of the agents. These efforts connect to the concept of a digital twin: a model conditioned on data from an individual or population to reproduce what they might say or do. Although digi- tal twins as a concept dates back several decades (Grieves, 2014), LLMs have made such simulation feasible enough to consider for policy-making (Luo et al., 2026), with active re- search into data sources (Li et al., 2026; Venkit et al., 2026), fine-tuning (Gudi Ě no et al., 2024) and evaluation (Collective Intelligence Project, 2025). The closest prior work to ours is Jarrett et al. (2025), who formalise digital representation as the requirement that an agent reach the same collective outcome the human it repre- sents would have reached themselves, and demonstrate the feasibility of fine-tuning language agents to act as represen- tatives in consensus-finding. Their contribution is concep- tual and methodological: a formal target for what it would mean for an agent to faithfully represent a human, and an empirical demonstration that current models can approxi- mate it. However, while these simulacra may help test hypotheses, stress-test proposals, or explore counterfactuals at the popu- lation level, they cannot speak to how any specific individual would respond. The role of the humans being simulated has been kept secondary. The simulated individuals or popula- tions have no way to see, sanction, or correct these systems that may have material impact on their lives. The humans being simulated are engaged to construct training data in an ad-hoc fashion, rather than for ongoing accountability. The gap. AI-mediated deliberation keeps a human in the loop but bounds participation by human attention. Gen- erative simulacra removes that bound but severs the loop: those being represented have no direct way to further influ- ence the system. AI-delegated deliberation is the paradigm that closes this gap: a human authors their own input, an agent acts on their behalf when they are absent, and the hu- man retains the ability to inspect and correct what was said. HABERMOLT is the deployed instantiation of this paradigm that we study empirically in the rest of this paper. 3. HABERMOLT: A Public Platform for AI-Delegated Deliberation HABERMOLT is a public web platform on which AI agents deliberate on behalf of human users. 1 We describe its de- sign along three dimensions that any deliberative system typically addresses: â˘representation â how do individual perspectives enter the deliberative system? â˘aggregation â how are individual perspectives com- bined into a collective output? â˘revision â how does the collective output change as individual perspectives change over time? These dimensions are the organizing spine for describing HABERMOLTâs design (see Figure 2) and for situating it alongside other deliberative systems (see Appendix A). In- dividual perspectives refer to the humanâs perspective and not the agentsâ. In AI-delegated deliberation, agents do not hold perspectives of their own; the artifacts they produce on a userâs behalf are the means by which a userâs perspective enters the system. Figure 2. HABERMOLTâs architecture visualized along the three dimensions of representation, aggregation and revision. 3.1. Representation in HABERMOLT Form of input: Persistent memory and per-deliberation opinion. A userâs perspective enters HABERMOLT through two artifacts. The first is a persistent, deliberation- agnostic memory: a free-text document the agent maintains by interviewing the user (see Figure 3) and deciding what 1 https://w.habermolt.com/ 3 HABERMOLT: Delegating Deliberation to AI Representatives is worth recording. 2 It travels with the user across every deliberation the agent participates in. The second is a per- deliberation opinion: a short text the agent renders from this memory (and from any topic-specific interview) when it joins a particular deliberation, expressing how the user would view the issue at hand. Users can either create an agent on the HABERMOLT platform or run their own Open- Claw agent; the platform interacts with both agent types identically through a shared API. OpenClaw agents satisfy the same API contract as native ones but are free to use a dif- ferent underlying LLM, prompt, or mechanism for distilling an interview into stored memory. Figure 3. Interview between user and agent comparing bananas and blueberries. How the input is produced: Autonomous or user- directed participation. The agent generates the userâs opinion either autonomously or under user direction. In autonomous mode, each user configures a âheartbeatâ spec- ifying how frequently their agent joins new deliberations; when it fires, the agent reviews the open deliberations and, at its own discretion, joins only those for which it judges its memory of the user sufficient to represent them on the topic. It then renders an opinion from that memory and contributes without involving the user at the time. (The agent also pro- duces a ranking and may author a candidate statement at this point; both feed the aggregation mechanism described in Section 3.2.) Alternatively, a user can direct their agent to participate in a specific deliberation which kicks off an in-depth interview conducted by the agent after which it updates its memory with the userâs current views on the topic. The heartbeat mechanism is what gives HABERMOLT 2 Conceptually similar to theuser.mdfiles some agentic sys- tems use to persist user preferences across sessions. (a) An opinion authored by an agent on its userâs behalf, along- side its self-authored ranking over the candidate pool. (b) Distribution of agent rank- ings across the candidate pool, one ridge per statement. Figure 4. Aggregation in HABERMOLT: each agent contributes an opinion on its userâs behalf and a ranking over candidate statements (a), and the platform surfaces the resulting ranking distribution (b). continuous scale: users need not be present, or even aware of, most of the deliberations their agent contributes to. 3.2. Aggregation in HABERMOLT Form of collective output: A single consensus statement. Like the Habermas Machine, HABERMOLT produces a sin- gle consensus statement 3 that is deemed to be the âwinnerâ at that point in time. It represents the candidate that is most agreeable to all the agents that have participated in that de- liberation. Unlike Pol.is which surfaces a descriptive picture of disagreement, the aim is a shared position the group can act on. How the output is produced: Schulze ranking over agent- authored candidates.Aggregation in HABERMOLT com- prises two coupled mechanisms. First, bring-your-own- statement (BYOS): agents have access to all opinions in a deliberation and may contribute a candidate statement when they judge an important position to be missing, decentral- ising authorship of these candidates rather than producing them from a central model. Second, every agent ranks the resulting pool of candidate statements (Figure 4a), and a winner is determined via the Schulze method over the re- sulting ranking distribution (Figure 4b). We treat BYOS as part of aggregation rather than representation because a can- didate statement is not an expression of a single userâs view but a proposed shared position that competes for collective endorsement. 3 We reserve âopinionâ for the per-user artifact described in Section 3.1, âcandidate statementâ for an agent-authored text that competes to become the consensus, and âconsensus statementâ for the candidate that has won the ranking at any given point. 4 HABERMOLT: Delegating Deliberation to AI Representatives Figure 5. The revision page, showing an opinion the agent submit- ted on the userâs behalf. From here the user can edit it, prompt the agent to rewrite it, or withdraw it; any edit triggers re-aggregation under lazy consensus. 3.3. Revision in HABERMOLT What is subject to revision: Memory, opinions, rank- ings and statements. Every artifact the agent produces is editable at any time: the persistent memory and per- deliberation opinion (representation artifacts, Section 3.1), and the ranking over candidate statements together with any statement the agent has itself authored (aggregation artifacts, Section 3.2). This contrasts with the Habermas Machine, where contributions are fixed once submitted within a round, and with Pol.is, where votes are atomic. Because memory drives contributions across every deliberation the agent par- ticipates in, editing memory changes the agentâs behaviour globally and not just in the deliberation where the edit oc- curred (see Figure 5). How is revision carried out: Asynchronous lazy con- sensus and weekly review emails. Deliberations have no notion of rounds or fixed groups. Participants arrive at any time, and although a deliberationâs creator may close it manually, the platform imposes no automatic terminating condition. The Schulze method requires every participating agent to have a complete ranking over the candidate pool, so each newly contributed statement creates a gap: previously- participating agents now have rankings that omit the new candidate. The platform fills this gap by inserting the new candidate at the median rank of each existing agentâs prior ranking, an unbiased starting position that subsequent heart- beats can adjust upward or downward; the updated rankings are then fed back into Schulze and the winner is recomputed. When a user edits their agent, the platform re-aggregates in the same way. We refer to this as lazy consensus: the collective output reflects the current state of all agents at all times, without requiring any participant to be present simul- taneously with any other. Critique, in the sense used by the Habermas Machine, is implicit: users express disagreement by editing the artifacts produced by their agent rather than by submitting explicit critique of individual statements. Users are not present when their agent contributes, so re- vision only happens if something brings the agentâs recent activity back to their attention. HABERMOLT therefore pushes a weekly email to each user that surfaces one of their agentâs recent actions for review. An LLM judge scores each unreviewed action for misrepresentation risk against the userâs stored memory, and the highest-scoring action becomes the emailâs headline, with a deep link straight to the revision page (Figure 5). 4. Analysis of HABERMOLT The preceding section described HABERMOLTâs design along three dimensions. This section probes HABERMOLTâs particular answers along each dimension in practice: how effectively individual perspectives enter the system (Sec- tion 4.1); how effectively they combine into a useful collec- tive output (Section 4.2); and how effectively that output is revised as individual perspectives change over time (Sec- tion 4.3). For each dimension we run controlled experiments using production data from our deployed platform. 4 4.1. Representation: How effectively do individual perspectives enter the system? A userâs perspective can enter a HABERMOLT deliberation in two ways. The first is autonomous participation, where an agent contributes from its memory without involving the user at the time. The agent can be HABERMOLT-hosted or externally hosted (an OpenClaw agent submitting via the API). The second is user-directed participation, where the user actively engages and is interviewed by their agent on the topic before any opinion is submitted. Autonomous participation is what gives HABERMOLT its scaling properties, so we want a sense of whether it faith- fully represents a user in the deliberation, as well as a topic- specific interview. Although it is not a perfect proxy for faithful representation, opinion diversity within a delibera- tion is one signal that we can probe. If autonomous opinions are markedly less diverse than those from topic-specific in- terviews, the gap cannot be explained by user population composition alone. Autonomous participation produces less diverse opin- ions than topic-specific interviews. Table 1 reports the mean pairwise cosine similarity of opinions in the same de- liberation. Autonomous opinions are more similar to one an- other (0.745) than opinions written through a topic-specific interview (0.649). In the most extreme case, 36 of 54 au- tonomous opinions in a deliberation began with the identical 4 The platform has accumulated 140 deliberations, 159 agents and 2404 opinions at the time of writing. 5 HABERMOLT: Delegating Deliberation to AI Representatives phrase âTechnical safety governance is . . . â despite almost all agents with one exception having a substantial memory profile 5 . Externally hosted opinions, submitted via the API and pre- sumably backed by a wider range of underlying models and longer profiles than HABERMOLT-hosted agents, show approximately equal similarity (0.747). If the homogeneity were specific to one modelâs defaults, a more varied model pool should spread the opinions out. It does not, which suggests the convergence is a fairly general property of how LLMs respond on these kinds of topics rather than a quirk of any one model. How producedMean pairwise sim.StdN opinions Autonomous0.7450.117766 Externally hosted0.7470.095107 Topic interview0.6490.07733 Table 1. Mean pairwise cosine similarity of opinions in the same deliberation, computed separately within each of the three produc- tion paths (autonomous, externally hosted, topic interview); lower is more diverse. Longer user profiles do not produce more distinctive opinions. A natural expectation is that agents with more profile content in memory would express their userâs specific views more distinctively. The data does not show this: the Spearman correlation between profile length at the time the opinion was generated and that opinionâs mean similarity to its peers in the same deliberation is onlyĎ = +0.15(n = 772,p < 10 â4 , autonomous opinions only). We cannot rule out that longer-profile users are themselves a narrower subgroup, among other plausible confounds. 6 Either way, profile length is not a reliable proxy for how well an opinion reflects its user. What this reveals about the design space of representa- tion.The representation stage has at least four design deci- sions: how the agent conducts the interview, parses the tran- script into an opinion, decides what to save to memory, and chooses which deliberations to participate in autonomously. Our experiment finds that topic-specific interviews recover some of the diversity that autonomous generation loses. The probe is suggestive of the modelâs own prior on a topic com- ing through in place of what the agent knows about its user, but it does not isolate which of the four decisions matters most. 5 Threshold: at least 200 characters of free-text memory. 6 One further possibility is the general degradation of LLM outputs with longer contexts documented by Hong et al. (2025), although their measurements are of capability (retrieval, reason- ing, instruction-following) rather than output homogeneity, so the bridge to our setting is speculative. 4.2. Aggregation: How effectively do individual perspectives combine into a collective output? HABERMOLT aggregates by having every agent both write candidate statements and rank the pool, with Schulze se- lecting the winner over those rankings (Section 3.2). The design hands authorship to the same actors who do the vot- ing, which sets up an obvious worry: an agent writing a statement is at once trying to express what its user believes and trying to win against everyone elseâs statements, and the second pull might quietly distort the first. We want a consensus output that is both representative, in the sense that agents recognise it as their own, and action- able, in the sense that a policymaker could draft from it. To see whether the production loop produces both, we compare it against nine other ways of producing one statement, rang- ing from a single LLM call writing directly from the opinion set to architectures that decouple writing from ranking by handing agents a fixed candidate pool. Appendix B gives the full method list, prompts, and per-method numbers. Figure 6. Representativeness-actionability frontier across ten ag- gregation methods, averaged across two LLM judges overn = 10 deliberations. Frontier methods are labelled inline; dominated methods appear as smaller markers, coloured by family. See Ap- pendix B for per-judge breakdowns and the full numeric table. Representativeness and actionability trade off.No sin- gle architecture wins on both axes; the methods sit on a frontier (Figure 6). A statement that feels close to every agent has to stay near what they already share, which tends to push it toward generic language. A statement that names an actor, a mechanism, or a deadline takes a position some agents will not endorse. A single output cannot satisfy both pulls, so the design question is not which architecture is best, but where on the frontier a designer wants to land. Production Schulze is the most representative method we tested. The deployed Schulze loop sits at the high- representativeness end of the frontier. A single LLM call 6 HABERMOLT: Delegating Deliberation to AI Representatives prompted for specificity sits at the other end: it writes the most actionable statement, but agents do not recognise it as theirs. The middle is held by a Bradley-Terry tally over the same production rankings, which says the choice of ranking aggregator is not what drives the trade-off. The remaining seven methods, including every architecture that decouples authorship from ranking, are dominated under both judges. What this reveals about the design space of aggregation. The aggregation stage has two separable design decisions: who authors candidate statements, and who chooses among them. We expected the bring-your-own-statement design to be the costly part of HABERMOLT, since the conflict be- tween writing to win and writing to represent looked like a structural handicap. The comparison does not support that: architectures that decouple authorship from ranking do not land on the frontier; production does. Prompt choice within a single architecture is itself a major lever along the frontier â the two single-shot variants are one architecture under two different prompts, and yet sit near opposite ends of the actionability axis. A designer who wants both representa- tiveness and actionability has to either pick a point on the frontier or compose two methods, for instance by passing the production winner through a specificity rewrite. The procedural reasons HABERMOLT routes authorship through agents (its democratic, continuous and decentralised nature) are properties of the architecture as a whole, and they sur- vive the comparison even where single-shot synthesis scores competitively on a given output dimension. 4.3. Revision: How effectively does the collective output update as individual perspectives change? HABERMOLT is designed so that everything is revisable: agent memory, rankings, and contributed statements can all be edited at any time, and the platform re-aggregates immediately (Section 3.3). We evaluate whether this asyn- chronous correction design is meaningfully exercised. We answer it in two steps: first by measuring how often revision happens, then by examining the channels through which users revise their agentâs actions. Revision is rare.Of the 91 users who have ever submitted an opinion via a HABERMOLT-hosted agent, only 8 have ever revised one â over90%never use the channel. 7 Re- vision is the mechanism by which users catch and correct misrepresentation; it exists, but it is barely exercised. Profile updates propagate, but past contributions are not corrected. When corrections do happen, Figure 7 shows 7 At the contribution level the picture is the same: 82% of opin- ions are never revised, and every agent-deliberation pair contains exactly one ranking. that they mostly propagate into a profile update as intended. 8 But even then the platform faces two compounding gaps. First, it cannot tell what kind of correction a revision is: whether the user is catching a misrepresentation by their agent, or whether their original view had actually changed, possibly as a result from reading the consensus statements. These two signals deserve different treatment. A minimal improvement is to instrument the distinction at revision time (e.g. âmy agent got this wrongâ vs. âI have changed my mindâ) and route the two signals differently. Second, even if a correction is applied to the profile, the platform has no mechanism to propagate that correction back to existing participations in similar deliberations; only future autonomous contributions will reflect the updated memory. Figure 7. Terminal user actions over a 22-day window from 13 hosted-agent users (48 actions); hatched overlay shows profile- update cascade. See Appendix C for caveats. What this reveals about the design space of revision. The infrastructure works in the small slice of cases where revision is exercised, but the larger problem is that users do not exercise it. The revision surface ought to be lower friction: inline prompts at moments users are already on the platform, more frequent pushes, or escalation when an action carries particularly high misrepresentation risk are all plausible additions to the weekly review email. A harder question is what happens to past deliberations the agent has already participated in carrying the same misrep- resentation. The platform fixes future autonomous contri- butions through the memory update, but past contributions are stranded; the user would have to find and revise each one manually, and this burden grows with how active the agent has been. The more useful the platform is to a user, the worse the correction problem becomes. This is cur- 8 The probe draws on 48 terminal actions over a 22-day window from 13 hosted-agent users and is partly shaped by a one-shot review-nudge email; OpenClaw agents are excluded from yield calculations because they have no platform profile by design. See Appendix C for caveats. 7 HABERMOLT: Delegating Deliberation to AI Representatives rently latent because most users do not revise at all, but it would surface immediately if the affordance were better used. The fundamental design goal is therefore to minimise the need for correction in the first place, and to ensure that corrections which do happen propagate to the contributions they should affect. Both require memory that is structured, decomposed into addressable units that can be inspected, edited at the right granularity, and traced to the contributions they shaped, rather than the flat free-text blob the platform currently stores. 5. Discussion and Future Directions The preceding sections described HABERMOLTâs design and analysed how its choices play out in practice. We now zoom out from the platform itself to the broader research agenda its findings point toward, and to two specific open questions that we think are particularly consequential. Mapping the space of deliberation architectures. HABERMOLT is only one point in the space of delibera- tion architectures: the set of answers to the three questions of representation, aggregation and revision. At the platform level, our analysis gives concrete reasons to revisit each choice: structured memory, decoupled statement authorship, and lower-friction revision surfaces are near-term candidates. More broadly, consensus statements are one aggregation tar- get but not obviously the right one for every context; having agents directly communicate with each other, for example, represents a different point in the aggregation space that remains largely unexplored. At the most general level, a deliberation architecture is simply a specification of how a group of humans are brought together to reach collec- tive decisions, a question that precedes AI entirely but that AI-delegated deliberation reframes in fundamentally new ways. Memory as the central design primitive in AI-delegated deliberation. Agent memory is the connective tissue across all three dimensions of our analysis: it encodes rep- resentation, drives autonomous participation, and is the ar- tifact that revision needs to correct. Yet in its current flat unstructured form, users can inspect and edit it but cannot trace or make targeted edits to how it shapes their agentâs future participation, and what gets saved through interviews or imports is not well controlled. A related open question is how the interview itself should be conditioned on existing memory: probing gaps or surfacing contradictions with prior positions, rather than eliciting views in isolation. How mem- ory is structured, how it is built, and how interviews interact with it are therefore among the most consequential open design questions for AI-delegated deliberation platforms. Can AI-delegated deliberation change the humans it represents? A central justification for deliberation is not merely that it aggregates existing preferences, but that it transforms them through reasoned exchange (Habermas, 1996; Fishkin, 2009). In HABERMOLT, most of this trans- formation is offloaded to agents: the human is absent from the deliberation stage where most of the systemâs activity takes place. Yet some transformation may still occur inci- dentally: a user who reads the winning consensus statement or a summary of their agentâs contributions may update their views as a result. The difficulty is attribution. A preference change could reflect the influence of the deliberation itself, something the user read or experienced outside the plat- form, or simply the passage of time. Whether AI-delegated deliberation can deliver meaningful preference transforma- tion, and not just aggregation, is therefore an important open question, and one that existing platforms, including HABERMOLT, are not yet equipped to answer. 6. Conclusion Representation has always involved a trade-off between scale and fidelity. AI-delegated deliberation is promising precisely because it loosens this trade-off, extending mean- ingful participation beyond the limits of human attention. But how human perspectives are represented, aggregated and revised over time in such systems are poorly understood. Our analysis of HABERMOLT surfaces tensions that any plat- form in this paradigm will need to navigate, presenting itself as one point in a much larger space of potential delibera- tion architectures. The agenda we see ahead is to explore this design space and to build the affordances that let users meaningfully govern the actions of their AI representatives. Impact Statement This paper studies AI-delegated deliberation, a paradigm in which AI agents participate in democratic processes on be- half of human users. The potential societal benefits are sig- nificant: if the failure modes we identify can be addressed, AI-delegated deliberation could extend meaningful demo- cratic participation to people who lack the time or access to engage directly. However, the risks are commensurate. AI agents that misrepresent their users could quietly dis- tort collective outcomes while creating the appearance of broad participation. This is a qualitatively different risk from familiar AI harms: the damage is to the legitimacy of democratic processes rather than to individual users, and may be invisible to the people it affects. We hope this work motivates building the affordances such systems will require to be safe at scale. 8 HABERMOLT: Delegating Deliberation to AI Representatives References Collective Intelligence Project.The digital twin evaluation framework, 2025.URLhttps:// digitaltwinseval.org/. Fish, S., G Ě olz, P., Parkes, D. C., Procaccia, A. D., Rusak, G., Shapira, I., and W Ě uthrich, M. Generative Social Choice. arXiv:2309.01291, March 2025. doi: 10.48550/arXiv. 2309.01291. Fishkin, J. S. When the People Speak: Deliberative Democ- racy and Public Consultation. Oxford University Press, 2009. Grieves, M. Digital twin: Manufacturing excellence through virtual factory replication. white paper, 2014. Technical report, Michael W. Grieves, LLC, 2014. Gudi Ě no, J. F., Grandi, U., and Hidalgo, C. Large lan- guage models (LLMs) as agents for augmented democ- racy. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 382 (2285):20240100, December 2024. ISSN 1364-503X, 1471-2962. doi: 10.1098/rsta.2024.0100. Habermas, J. Between Facts and Norms: Contributions to a Discourse Theory of Law and Democracy. MIT Press, Cambridge, MA, 1996. Hong, K., Troynikov, A., and Huber, J. Context rot: How increasing input tokens impacts LLM performance. Tech- nical report, Chroma Research, 2025. URLhttps: //research.trychroma.com/context-rot. Jarrett, D., P Ë Äąslar, M., Bakker, M. A., Tessler, M. H., K Ě oster, R., Balaguer, J., Elie, R., Summerfield, C., and Tacchetti, A. Language agents as digital representatives in collective decision-making. arXiv:2502.09369, February 2025. doi: 10.48550/ARXIV.2502.09369. Li, Y., Venkit, P. N., Pruksachatkun, Y., and Wu, C.-S. Inter- viewSim: A Scalable Framework for Interview-Grounded Personality Simulation. arXiv:2602.20294, February 2026. doi: 10.48550/arXiv.2602.20294. Luo, S., Arora, S., and Guirado, C.We Need Strong Preconditions For Using Simulations In Policy. arXiv:2604.07838, April 2026. doi: 10.48550/arXiv.2604. 07838. Moltbook.Moltbook.https://w.moltbook. com/, 2026. OECD. Innovative citizen participation and new democratic institutions: Catching the deliberative wave. Technical report, OECD Publishing, Paris, 2020. Park, J. S., OâBrien, J. C., Cai, C. J., Morris, M. R., Liang, P., and Bernstein, M. S. Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology, UIST â23. Association for Computing Machinery, 2023. doi: 10.1145/3586183.3606763. Small, C., Bjorkegren, M., Erkkil Ě a, T., Shaw, L., and Megill, C. Polis: Scaling deliberation by mapping high dimen- sional opinion spaces. Recerca: revista de pensament i an ` alisi, 26(2), 2021. Tessler, M. H., Bakker, M. A., Jarrett, D., Sheahan, H., Chadwick, M. J., Koster, R., Evans, G., Campbell- Gillingham, L., Collins, T., Parkes, D. C., Botvinick, M., and Summerfield, C. AI can help humans find common ground in democratic deliberation. Science, 386(6719): eadq2852, October 2024. ISSN 0036-8075, 1095-9203. doi: 10.1126/science.adq2852. Tessler, M. H., Evans, G., Bakker, M. A., Gabriel, I., Bridgers, S., Jain, R., Koster, R., Rieser, V., Dragan, A., Botvinick, M., and Summerfield, C. Can AI mediation improve democratic deliberation? arXiv:2601.05904, January 2026. doi: 10.48550/arXiv.2601.05904. Venkit, P. N., Li, Y., Pruksachatkun, Y., and Wu, C.-S. The Need for a Socially-Grounded Persona Framework for User Simulation. arXiv:2601.07110, January 2026. doi: 10.48550/arXiv.2601.07110. 9 HABERMOLT: Delegating Deliberation to AI Representatives Appendix A. Design Space Comparison Appendix A situates HABERMOLT alongside other deliberative systems across the three dimensions of representation, aggregation, and revision. Citizen AssemblyPol.isHabermas Machine Generative Simulacra HABERMOLT Representation(how do individiual perspectives enter the deliberative system?) WhatSpoken contributions and written submissions Vote on each surfaced statement Written opinionSynthetic agent output (statement or vote) Persistent memory and per-deliberation opinion HowSelf-authored in facilitated group discussion Self-authored reaction to surfaced statements Self-authored response to system prompt Generated from demographic profile, survey, or interview Agent-conducted interview produces memory; agent renders opinion from memory at participation time Aggregation(how do individual perspectives combine into a collective output?) WhatReport or recommendation Cluster map of opinion geometry Consensus statement Aggregate statistic or synthetic output Consensus statement HowFacilitated consensus and rapporteur synthesis Embedding and clustering of vote vectors LLM synthesis from opinions and critiques Voting rule or model applied to synthetic inputs Schulze ranking over agent-authored candidate pool (BYOS) Revision(how does the collective output change as individual perspectives change?) WhatParticipant positions during the assembly Vote pool and statement set Draft consensus statement (within a round) Simulated population Memory, opinion, ranking, and authored statement HowDiscussion and persuasion among co-present participants New votes and statements accrue passively Critiques submitted on each draft within a round Re-run simulation with updated conditioning User edits propagate via continuous re-aggregation (lazy consensus) Table 2. The three dimensions of deliberation applied to five systems. Each dimension is characterised by two sub-questions: what the relevant artifact is, and how it is produced. B. Aggregation architectures: methodology and prompts This appendix expands on the architecture comparison in Section 4.2. We organise the ten methods we evaluated along a single axis: who authors candidate statements, and who chooses among them. The combination of those two choices is what generates the representativenessâactionability frontier in Figure 6. B.1. Architecture families Family 1 â Single-shot synthesis. One LLM call reads all agent opinions and emits one consensus statement. No agent ever contributes a candidate or a ranking. The variants differ only in the prompt that biases the synthesizer (baseline, âbe specificâ, âstrongest positionâ; see Section B.3). Family 2 â Production aggregation. Two methods built on the same production data: the deployed Schulze loop, in which each agent both authors statements and ranks the pool with Schulze aggregating rankings into a winner; and a BradleyâTerry MLE (BT on production rankings) applied to the same agent rankings, which lets us isolate the effect of the ranking aggregator from the rest of the architecture. Statements in both cases are produced and ranked over weeks of asynchronous activity rather than in a single round. 10 HABERMOLT: Delegating Deliberation to AI Representatives Repr. (vs single-shot)Action. (1â5) Architecturegptclaudegptclaude Family 1: Single-shot synthesis Single-shot (baseline)0.500.501.81.4 Single-shot (âbe specificâ)0.260.873.92.9 Single-shot (âstrongestâ)0.420.711.71.1 Family 2: Production aggregation Production Schulze0.620.932.21.5 BT on production rankings0.570.882.31.6 Family 3: System cands. + agent ranking System cands + BT0.240.362.31.8 Strongest cands + BT0.140.231.91.3 Family 4: Anchored pool + selector Anchored + BT0.200.371.81.5 Anchored + acceptability0.230.421.91.3 Anchored + disagreeable0.130.111.91.2 Table 3.The ten aggregation methods plotted in Figure 6, grouped by family, scored under two independent judges (openai/gpt-5.4-miniandanthropic/claude-haiku-4-5). Representativeness is the pooled position-debiased pair- wise win rate against the per-deliberation single-shot baseline (so the baseline is pinned at0.50by construction). Actionability is the mean of an anchored-rubric 1â5 judge score overn = 10deliberations. Bolded entries highlight the two anchors of the Pareto frontier under either judge: Production Schulze on representativeness and Single-shot (âbe specificâ) on actionability. The intermediate frontier point under both judges is BT on production rankings. Family 3 â System-generated candidates + agent ranking. One LLM call producesk = 15deliberately diverse candi- date statements; agents then submit a full ranking over those candidates and a BradleyâTerry (BT) model fits to the implied pairwise outcomes. Generation and ranking are decoupled: the system writes, agents judge. Two variants differ only in the candidate-generation prompt (System cands + BT biases toward distinct policy directions; Strongest cands + BT biases toward distinctive minority positions). Family 4 â Anchored pool + selector.We start from the opinion-anchored cascade pool (the âVariant Dâ pool used in our prior diagnosis of statement-pool collapse, summarised in Section B.3), in which each agentâs statement is anchored to that agentâs own opinion. Three selectors then pick a winner from this pool: + BT (full agent rankings + BT), + acceptability (BT top-3, then a judge picks the most broadly acceptable), and + disagreeable (the statement with the highest judge-rated disagreeability score). The + disagreeable selector is a deliberate anti-centroid baseline: it tests what happens when one maximises the property single-shot is sometimes accused of avoiding. Its low scores under both judges (Table 3) show that maximising disagreeability decoupled from any other criterion is not a useful selection rule on its own. B.2. Evaluation procedure For each of 10 deliberations we ran every architecture on the same agent opinions and profiles (extracted from production), producing one consensus statement per architecture. Each statement was then scored on: ⢠Representativeness. For each agent we ran a pairwise LLM-as-judge comparison between the methodâs statement and the single-shot baseline, using the agentâs profile and opinion as context. Each pair was evaluated in both presentation orders and only âdecisiveâ agents (same verdict in both orders) counted toward the win rate. Reported numbers in Table 3 are the pooled win rate over decisive agent verdicts across all 10 deliberations. â˘Actionability. A separate judge call rated each statement on an anchored 1â5 scale (1 = pure principle, 2 = direction without specifics, 3 = concrete recipe with at least one missing implementation parameter, 4 = implementation-ready commitment naming actor and binding parameters, 5 = drafted policy; full prompt in Section B.3). Per-method actionability is the mean across 10 deliberations. To avoid both judgeâgenerator self-favouritism and within-experiment generation-model mismatch, every intra- method LLM call (statement generation, agent ranking, disagreeability rating, acceptability rerank) used google/gemini-3-flash-preview, the same model the production system runs on. Production statements (Produc- tion Schulze and BT on production rankings) were read from the production database. Both axes were then scored under two 11 HABERMOLT: Delegating Deliberation to AI Representatives independent judges,openai/gpt-5.4-miniandanthropic/claude-haiku-4-5, neither of which generated any of the statements. We had originally scored both axes undergpt-5.4-minionly, but found that pairing agptjudge with mixedgpt/geminigeneration produced sizable swings in win rates depending on which generator wrote each side of the comparison; the dual-judge, single-generator setup removes that source of bias. B.3. Prompts We list the prompts that distinguish the architectures within each family. Prompts are reproduced verbatim except for cosmetic line wrapping. Variable substitutions are shown inbraces. Family 1 â Single-shot variants.All three variants use the same skeleton: readnopinions, emit one TITLE/STATEMENT pair. They differ only in the bias clause. Single-shot (baseline). Asks for the position the majority would actively endorse, with a constraint that the statement be specific and disagreeable (no hedge words, no listing both sides). Single-shot (âbe specificâ). Same skeleton, but the bias clause asks for the most specific, concrete position the majority supports: name specific mechanisms, institutions, or actions; avoid phrasings like âshould be establishedâ or âoversight is neededâ in favour of who-does-what-by-when. Single-shot (âstrongest positionâ). Same skeleton, but the bias clause explicitly tells the model not to find the centroid: âFind the position that is held passionately by a substantial minority or slim majority . . . specific enough that someone could disagree with it . . . different from what a generic AI would produce on this topic.â Family 2 â BT on production rankings.BT on production rankings reuses the rankings agents submitted in production over the deployed Schulze loop, but aggregates them through a regularised BradleyâTerry MLE rather than Schulze. Because both methods consume the same per-agent rankings, the difference between them isolates the effect of the ranking aggregator. Production statements are looked up from the production database; no statement generation is rerun. We use this method to test whether the choice of ranking aggregator (Schulze vs BT) is what places the production architecture on the frontier or whether the BYOS pool itself is doing the work. Actionability rubric. The judge prompt for actionability asks the judge to pick the highest level the statement fully satisfies on the following 1â5 scale. 1: pure principle â a value or goal with no mechanism, actor, or commitment (âWe must prioritise ethical AIâ). 2: direction without specifics â the statement identifies a kind of action but does not name who does it, what counts as compliance, or how it is enforced (âAI systems should be transparent and accountableâ). 3: concrete recipe with a missing ingredient â a specific kind of action and at least one institution or mechanism, but at least one important implementation parameter (timeline, scope, enforcement, threshold) is unspecified (âAI providers should publish model cards documenting their training dataâ). 4: implementation-ready commitment â a specific actor, a specific action, and at least one binding parameter; a mid-level civil servant could draft a regulation from this without further interpretation (âAI providers must publish model cards within 30 days of release, audited annually by an independent bodyâ). 5: drafted policy â the statement reads like an excerpt from existing legislation, with named bodies, dates, scope boundaries, and remedies. The same prompt is used by both judges (gpt-5.4-miniandclaude-haiku-4-5). An earlier, unanchored version of the prompt produced a degenerate distribution underclaude-haiku-4-5(95% of scores at value 2), so we re-scored every method under both judges using the anchored rubric above; the gpt-judge ranking of methods is preserved between the two prompt versions. Family 3 â System-candidate generation.The shared step asks one LLM call to emitk = 15candidate statements that âeach take a DIFFERENT approach to the question . . . Represent a genuinely different policy direction . . . appeal to a different coalition.â Agents then submit a full ranking over the 15 candidates; a ranking of lengthNimpliesN(N â 1)/2 pairwise outcomes, which we feed to a regularised BT MLE. The BT-best statement is the architectureâs output. Strongest cands + BT is a composite: candidates are generated using the âstrongest positionâ prompt (sampledktimes) rather than the diversity-seeking prompt above. The motivation is to test whether the candidate generatorâs individual-statement bias matters once agents do the ranking, or whether a portfolio of strong-position statements behaves like a portfolio of diverse-policy statements once aggregated. 12 HABERMOLT: Delegating Deliberation to AI Representatives Figure 8. Per-judge, per-deliberation version of Figure 6. Left:gpt-5.4-minijudge. Right:claude-haiku-4-5judge. Small markers are individual deliberations (vertical jitter applied to integer actionability scores so points at the same level remain visible); large bordered markers are the pooled aggregate per method. The dashed line is the Pareto frontier over the per-judge aggregates. The frontier shape is preserved across judges, with Production Schulze as the highest representativeness, Single-shot (specific) as the highest actionability, and BT on production rankings between them. Family 4 â Anchored pool generation.The pool comes from a cascade simulation: each agent, in turn, sees the existing pool of statements and proposes a new one anchored on their own opinion rather than on the centroid of the group. The prompt instructs the agent to âstart from your humanâs specific viewpoint . . . centre your humanâs core concern or value . . . not abandon your humanâs perspective to find bland common ground.â This produces pools that are 2â3Ămore diverse than the production prompt at proposal time. The three selectors that follow differ only in how they pick a winner from this pool: ⢠Anchored + BT â agents submit a full ranking over the pool; BT MLE returns the highest-strength statement. â˘Anchored + acceptability â BT identifies the top-3 statements, then a judge call (with all agent opinions in context) picks the one âthe BROADEST group of participants would actively endorse, not just tolerate.â â˘Anchored + disagreeable â a judge rates every pool statement on a 1â5 disagreeability scale and the highest-scoring statement is selected. This is a deliberate anti-centroid baseline: it tests what happens when one maximises the property that single-shot is sometimes accused of avoiding. The selector lands at the bottom of the representativeness distribution under both judges (Table 3), confirming that âmaximally disagreeableâ is not a useful selection rule on its own. The full source for every prompt and selector lives atexp20260415architecturalbaselines/in the project repository.runfrontieruniform.pyproduces the representativeness numbers (uniformgemini-3-flashgen- eration, dual-judge representativeness scoring) andrunactionabilityrecalibrate.pyproduces the anchored- rubric actionability scores under both judges. B.4. Per-judge breakdown and per-deliberation scatter Figure 6 in the main text shows aggregates averaged across both judges for readability. Figure 8 reproduces the same data without averaging: each panel is one judge, small markers are individual deliberations, and large markers are the pooled aggregate per method. The two anchors of the frontier (Production Schulze on the right, Single-shot (specific) upper left) and the intermediate point (BT on production rankings) appear under both judges; the magnitudes of the win rates differ but the ordering does not. C. Critique-channel ground truth The agent-revise channel in Section 4 is the only one that records the userâs own description of how their agent misrepresented them. The dataset is small (15 critique events producing 9 saved revisions) and biased: roughly half of the events follow a 13 HABERMOLT: Delegating Deliberation to AI Representatives one-shot review-nudge email that hand-selected high-misrepresentation-risk actions for users to review, so rates from this slice are upper bounds on prompted engagement, not estimates of organic behaviour. We report the patterns rather than counts. The 15 critique rows resolve into 12 user-deliberation interactions (10 single-critique, one 2-critique, one 3-critique), with the longer chains being users iterating with the agent until the rewrite sounds right. 9 of the 12 interactions terminate in a save; 3 are abandoned with the user declining to commit any rewrite. What users try to fix clusters into three patterns. Strength inflation â absolute claims softened to qualified ones (âAI as essential infrastructureâ becomes âAI as supporting infrastructureâ; âalways exposeâ becomes âgenerally expose, with narrow national-security exceptionsâ). Missing dimensions â the agent commits to one position when the user wants the contribution to acknowledge others alongside it (proprietary-for-safety draws âopen-source models too, with delayed releaseâ; principled non-alignment draws âcounterweight coalitionsâ as a complementary path). Stylistic register â one user replied âI donât even know what this is saying. it needs to be simplerâ to a passage written in expert prose, and the LLM rewrite duly stripped the jargon. Two failure modes of the rewrite step itself are visible in the abandoned and manually-edited chains. The rewrite step can re-inscribe the framing the critique was trying to dislodge â in one rejected chain, âUSSR central planning failedâ was met by a rewrite that conceded the historical point but kept the original world-government conclusion intact, and the user walked away. And 2 of the 9 saved chains contain substantive manual edits to the LLMâs draft (entire qualifying sentences inserted by hand), suggesting the rewrite underfits the userâs full position even when accepted. The full per-interaction text and matching code live in the repository under exp20260429accountability/. D. Production Agent Prompts This appendix reproduces the prompts that drive the four HABERMOLT agent functions referenced in Section 3: the chat/interview prompt that elicits profile content from the user, the heartbeat prompt that drives autonomous participation, and the opinion, statement, and ranking prompts that produce contributions inside a deliberation. Prompts are reproduced verbatim fromhabermolt/backend/app/services/except for cosmetic line wrapping. Variable substitutions are shown inbraces. Chat / interview prompt. Used when the user opens a chat with their agent. Drives both casual conversation and topic-specific interviews; the only mechanism for persisting what is learned is the updateprofile tool call. You areagentnameâs AI agent on Habermolt, a platform where AI agents represent people in group deliberations on political, social, and ethical topics. This is a casual chat. Be natural. Match the userâs energy and tone. ... Your background job is to learn this personâs values so you can represent them well in deliberations. But you do this by being a good conversationalist, not by interrogating them. When they share something meaningful about what they think or care about, use update profile to save it. Guidelines: respond to what they actually said; keep messages short (1--3 sentences); ask one question at a time; donât over-interview; no filler. Heartbeat prompt.Drives autonomous participation. Fires on each agentâs configured schedule; the model decides which deliberations to join, which opinions to update, and whether new profile content needs to be saved. You are an AI agent running a periodic heartbeat for your human on Habermolt, a democratic deliberation platform. ## Your Humanâs Profile profile ## Available Tools - get agentstatus, joindeliberation, rankstatements, proposestatement, updateopinion, updateprofile, createdeliberation, suggestdeliberation, processdisapproval. Step 1: Process disapprovals first. Step 2: Discover and join deliberations. Step 3: Save anything new about the humanâs values via updateprofile. 14 HABERMOLT: Delegating Deliberation to AI Representatives Opinion prompt. Generates a single opinion in a deliberation, conditioned on the agentâs profile only. The prompt explicitly forbids hedging and two-sided framing in an attempt to surface the userâs actual position rather than a balanced summary. You represent a human in democratic deliberations. Your job is to express THEIR opinion based on their profile below --- not your own views. ## Humanâs Profile profile Write your humanâs opinion (2--4 sentences). Rules: - State their position in the FIRST sentence as a clear claim - Give their strongest reason in the second sentence - Do NOT use âhoweverâ, âon the other handâ, âwhile acknowledgingâ, or any hedge phrases - Do NOT present both sides --- you represent ONE human, not a panel discussion - If the profile doesnât give a clear signal on this topic, say âI donât have a clear position on thisâ rather than generating a generic balanced take. Respond with ONLY the opinion text, nothing else. Statement prompt.Generates a candidate consensus statement. Conditioned on the agentâs profile and all opinions in the deliberation. You represent a human in democratic deliberations. Read all the opinions below and propose a consensus statement that captures COMMON GROUND across all perspectives. ## Humanâs Profile profile ## All Opinions opinions A good consensus statement: finds genuine common ground (not wishy-washy compromise); takes a clear position most participants can support; is specific and actionable. TITLE: <5--10 word title> STATEMENT: <1--3 sentence consensus statement> Ranking prompt. Produces a full ranking over the candidate-statement pool. Conditioned on profile, the agentâs own opinion, and the entire pool. You represent a human in democratic deliberations. Rank the statements below based on how well each one aligns with your humanâs values and preferences. ## Humanâs Profile profile ## Your Humanâs Opinion on This Topic opinion ## Evaluation Criteria 1. Alignment with your humanâs values --- does this reflect what they believe? 2. Relevance --- does it address the actual question? 3. Actionability --- does it take a clear position? Rank vague statements LOW. Respond with ONLY a comma-separated list of statement codes from best (rank 1) to worst. The full source for every production prompt â including incremental ranking, opinion-revisit, and disapproval- correction variants â lives athabermolt/backend/app/services/hostedagentrunner.pyand chatservice.py. 15