Paper deep dive
An Evidence-Grounded Multi-Agent System for High-Level Bio-Robot Design
Yujun Chen, Tianle Li, Jiayu Chen, Zhen Yin
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 8/21/2026, 2:50:22 AM
Summary
The paper introduces micro_biorobot_agent, an offline multi-agent system built on Qwen3.5-27B for high-level bio-robot design. It utilizes a shared workspace to coordinate specialized agents for requirement analysis, module-specific retrieval, candidate assembly, conflict checking, and validation over an integrated library of 23,762 records. The system employs rule-based output checks to correct false gaps, unsupported part mentions, and source-tracking errors, achieving high scores in evaluation sets compared to baseline systems.
Entities (10)
Relation Signals (8)
micro_biorobot_agent → uses → Qwen3.5-27B
confidence 99% · We present micro_biorobot_agent, an offline multi-agent system built on Qwen3.5-27B.
micro_biorobot_agent → utilizes → Integrated Evidence Library
confidence 95% · The system combines ... over an integrated library of 23,762 records
Integrated Evidence Library → includes → iGEM
confidence 92% · iGEM provides engineered parts with DNA sequences
Integrated Evidence Library → includes → FPbase
confidence 92% · FPbase provides fluorescent proteins
micro_biorobot_agent → corrects → false gap
confidence 90% · Deterministic output checks ... correct false gaps
micro_biorobot_agent → corrects → source-tracking error
confidence 90% · Deterministic output checks ... correct ... source-tracking errors.
Integrated Evidence Library → includes → Kosuri 2013
confidence 90% · Kosuri 2013 provides measured promoter and RBS performance
Cello → compiles → DNA circuits
confidence 85% · Cello compiles formal logic specifications into DNA circuits
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:In this paper, a bio-robot is an engineered living or biohybrid system in which living cells perform one or more core functions, such as sensing, information processing, actuation or output. We focus on systems whose cell-based functions are programmed by genetic circuits; physical movement is optional. Designing such a system requires translating application requirements into sensing, logic or memory, output, assembly, host and containment modules, while grounding each choice in traceable parts and evidence. We present micro_biorobot_agent, an offline multi-agent system built on Qwen3.5-27B. The system combines requirement analysis, module-specific retrieval, candidate assembly, conflict checking, local repair, independent review and validation over an integrated library of 23,762 records covering biological parts, measured combinations, literature-supported relationships and actuation evidence. Deterministic output checks align the final report with the retrieved part set and correct false gaps, unsupported part mentions and source-tracking errors. On two author-developed evaluation sets of 50 queries each, the system obtains mean overall scores of 7.35 and 8.04, the highest among the seven evaluated systems; on Scenario Design it exceeds the runner-up by 2.23 points. A 50-query paired ablation shows that the source-tracking check reduces false-gap incidents from 15 to 3, an 80% reduction, and increases source accuracy by 0.75 points. This paper reports the Qwen3.5-based v1 system and evaluates high-level design reports rather than experimentally validated circuits.
Tags
Links
- Source: https://arxiv.org/abs/2608.19699v1
- Canonical: https://arxiv.org/abs/2608.19699v1
Trouble viewing inline? Open PDF directly →
Full Text
52,206 characters extracted from source content.
Expand or collapse full text
An Evidence-Grounded Multi-Agent System for High-Level Bio-Robot Design Yujun Chen Tianle Li Jiayu Chen Zhen Yin Abstract In this paper, a bio-robot is an engineered living or biohybrid system in which living cells perform one or more core functions, such as sensing, information processing, actuation or output. We focus on systems whose cell-based functions are programmed by genetic circuits; physical movement is optional. Designing such a system requires translating application requirements into sensing, logic or memory, output, assembly, host and containment modules, while grounding each choice in traceable parts and evidence. We present micro_biorobot_agent, an offline multi-agent system built on Qwen3.5-27B. The system combines requirement analysis, module-specific retrieval, candidate assembly, conflict checking, local repair, independent review and validation over an integrated library of 23,762 records covering biological parts, measured combinations, literature-supported relationships and actuation evidence. Deterministic output checks align the final report with the retrieved part set and correct false gaps, unsupported part mentions and source-tracking errors. On two author-developed evaluation sets of 50 queries each, the system obtains mean overall scores of 7.35 and 8.04, the highest among the seven evaluated systems; on Scenario Design it exceeds the runner-up by 2.23 points. A 50-query paired ablation shows that the source-tracking check reduces false-gap incidents from 15 to 3 (−80%-80\%) and increases source accuracy by +0.75+0.75. This paper reports the Qwen3.5-based v1 system and evaluates high-level design reports rather than experimentally validated circuits. 1 Introduction Synthetic biology seeks to make biological systems easier to compose, reuse and engineer [9, 5]. Biohybrid microrobots are commonly described as living organisms integrated with artificial carriers for sensing, actuation and control [1]. We use bio-robot as a broader operational term for an engineered living or biohybrid system in which cells perform one or more core functions, including sensing, information processing, memory, actuation or output. In this paper, bio-robot design refers to the genetic-circuit layer in the living component; movement or an artificial carrier may be present, but is not required. Such systems can release payloads in tumour microenvironments [8] and can be constrained by host limitation and kill-switches [6]. The same design logic extends to gut living medicines, infection and wound monitoring, pollution sensing, bioremediation and on-demand biomanufacturing. Their value lies in the fact that living cells operate naturally in wet, chemically rich and spatially heterogeneous settings where silicon hardware is difficult to deploy for long periods. Design is a major bottleneck. A short request such as “a bacterial micro-robot that is activated in hypoxic tumour tissue and records its arrival” already implies a sensing module, a logic or memory module, an output reporter, an assembly scaffold, a host or vector choice and a containment module. The designer must translate the scenario into functional modules, search iGEM [12], FPbase [17], the Synthetic Biology Open Language (SBOL) and SynBioHub [11, 21], papers and local tables for real parts, and then check sequences, interfaces, host range, evidence and missing-part boundaries. Omitting one prerequisite module makes the design incomplete; mismatching one part can make the circuit non-executable; and claiming a part that is not in the library undermines trust. Bio-robot circuit design is therefore a system-design problem rather than a single-part lookup. A useful design platform must infer unstated goals and prerequisite modules, choose candidates with explicit evidence and interface information, state gaps accurately and explain the design well enough for an engineer to review. We evaluate these requirements through completeness, compatibility, design detail, source accuracy and clarity. Instead of passing retrieval results to one large language model (LLM), micro_biorobot_agent uses a shared structured workspace, traditionally called a blackboard. Approximately two dozen specialized components, including LLM agents, deterministic rules and validators, handle specification, scenario interpretation, module-specific retrieval, candidate design, conflict resolution, repair and error review. Their source-tagged reports are written to a shared structured design record so a report can be assembled, challenged and repaired rather than emitted in one pass. Retrieval-augmented LLMs are attractive for this task, but full designs expose three trust failures that are damaging because they often appear plausible. In a false gap, the model states that no library part covers a module even though a suitable part was retrieved. In an unsupported part mention, it names a real library card that was not retrieved as if it were available. In a source-tracking error, it marks a part that appears in the recorded retrieved-part list as out of library or not retrieved. These errors hide usable solutions, hallucinate availability and cause the system to deny its own evidence. We correct them with rule-based output checks rather than prompt engineering, leaving the retriever’s ranking metrics structurally unchanged. The contributions are: • A shared-workspace workflow (Sec. 3.1) that coordinates specialized LLM agents, deterministic rules, candidate generation, conflict checking, repair, error review and validation through a shared structured design record. • An integrated evidence library (Sec. 3.2): 23,762 records comprising 6,009 part-like entries, 17,715 measured-combination or relationship records, and 38 physical-actuation method/evidence records, with a separation rule that prevents bulk data from reducing search quality. • Rule-based output checks (Sec. 3.6) that correct false gaps, label unsupported part mentions and repair source-tracking errors without changing candidate IDs or retrieval ranking. • Five evaluation criteria and a common scoring procedure (Sec. 4), under which the v1 system ranks first on both 50-query subsets, and a paired output-check ablation reduces false gaps by 80%. Study scope. This paper reports v1 with Qwen3.5-27B [26]. Performance comparisons refer to the task sets, evaluation criteria and seven systems evaluated in this study. The system generates high-level candidate reports for expert review. It runs offline end-to-end with a CPU fallback (BM25 retrieval and a Python standard-library client to a local model endpoint) and an offline-GPU path (dense vector search and a locally served Qwen LLM). 2 Related Work Retrieval-augmented part recommendation. Retrieval-augmented generation (RAG) grounds LLM output in a corpus [19]. On the retrieval side, sparse BM25 and dense vector indexes provide interpretable lexical matching and scalable nearest-neighbour search [27, 14]. We use this backbone, but grounding alone does not ensure consistency: an answer can cite a retrieved part and elsewhere claim that the same module is empty. Our output checks close this gap deterministically rather than relying on the model to remain self-consistent. Synthetic-biology part libraries. Standardised parts and design-exchange languages provide the substrate for automated design [5, 11, 21]. Repositories such as the iGEM Registry [12], the Kosuri promoter×ribosome-binding site (RBS) composability library [15], and FPbase [17] provide real sequences or verifiable entries; the Kosuri and Mutalik resources also provide measured expression elements and pairwise composability evidence [22, 15]. We integrate these sources with evidence linked by digital object identifiers (DOIs) under one record format. A separate-index rule prevents large auxiliary datasets from reducing the quality of the default search results. Genetic-circuit design automation. Cello compiles formal logic specifications into DNA circuits using characterised gate libraries and validates the resulting designs by simulation and experiment [23]. Our task is complementary: it starts from open-ended application requests and produces evidence-bounded candidate reports rather than Cello-compiled or experimentally validated DNA constructs. The design-detail score used here must therefore not be read as proof of biological function. Multi-agent and tool-using reasoning. Shared-workspace systems, often called blackboard systems, were introduced to coordinate local experts under uncertainty [10]. Recent LLM systems extend this idea through reasoning and acting loops, tool use and multi-agent conversation [33, 29, 32]. Our system stores all agent outputs in a structured design record with explicit source links, making the process easier to inspect and test. Scientific and biological agents. Coscientist and ChemCrow combine LLMs with scientific tools and automation [3, 4]; the Virtual Lab uses a team of agents for nanobody design followed by experimental validation [30]. CRISPR-GPT integrates planning, retrieval and specialised tools for gene-editing workflows [25], whereas GeneAgent uses biological databases for self-verification [31]. Our system focuses on evidence organization, consistency with retrieved parts and modular design reports for synthetic-biology applications. Evaluation of biological AI systems. LAB-Bench evaluates practical biology-research capabilities against expert references [18]. Our benchmark instead targets open-ended, module-level bio-robot design and evaluates the structure, evidence use and consistency of generated design reports. Model-based evaluation. Model-based scoring [34] scales qualitative assessment, but scores can vary and may favour the evaluator’s own response style. Work on factuality and hallucination also shows that fluent answers are not necessarily evidence-grounded answers [20, 13]. We use the same scoring instructions for all systems and follow model-based scoring with a separate source check. Paired comparisons use the same queries and scoring instructions before and after a component is changed. Rule-based output checks. In contrast to GeneAgent’s model-driven database self-verification [31], our checks compare exact part identifiers. One check labels parts that were not retrieved, and another removes an incorrect label from parts that were retrieved. These checks verify consistency with the search results rather than biological function. 3 System and Method micro_biorobot_agent combines a shared multi-agent workspace with an integrated evidence library. Retrieved parts are organized by biological module, reviewed for conflicts and source support, and then rendered as a design report. Rule-based checks compare the final report with the retrieved part list before it is returned to the user. 3.1 Multi-Agent Workflow and Shared Design Record The central design principle is that no single agent or stage owns the full design. Approximately two dozen specialized components collaborate through a shared design record (Table 1). It stores the user requirements, required modules, candidate parts, interface checks, evidence references, review findings, repair history and the identifiers of retrieved parts. Each stage writes a source-tagged report to this record, allowing the final report to be traced back to the component and evidence behind each claim. Figure 1 summarizes the architecture. A supervisor routes work among five specialist agent groups, all of which read and write a shared design state. The evidence library supplies parts, measurements, literature relationships and actuation evidence. Deterministic output checks reconcile the rendered answer with the retrieved part set before the final report is returned. Figure 1: Multi-agent architecture of micro_biorobot_agent. Specialist agent groups share a structured design state while a supervisor routes work. The evidence library supplies grounded records, and deterministic output checks align the final report with the retrieved part set. Two-headed arrows denote shared-state read/write access; single-headed arrows denote data or control flow. Table 1: Main stages of the multi-agent workflow. stage components purpose requirements analysis agents and supervisor identify the goal, constraints and required modules evidence search scenario and evidence agents gather design patterns, evidence and safety needs part retrieval module-specific retrieval agent find candidate parts for each required module design design agents assemble several internally consistent candidates review and repair conflict checker, repair planner and three review agents check interfaces, biological mechanisms and evidence; revise affected modules validation seven rule-based validators check evidence, completeness, compatibility, feasibility, safety, unsupported claims and output format Independent error review. Before the report is finalized, three review agents examine the proposed design. They check whether adjacent modules can connect, whether each part has the required biological behaviour, and whether each claim is supported by a library record or paper. Every finding must refer to an existing part or evidence ID; unsupported findings are discarded. Validation and focused repair. Seven rule-based validators check evidence, module coverage, interfaces, physical feasibility, safety, unsupported claims and output format. A report can pass, receive a warning, require repair or be blocked for safety. When possible, only the affected module is revised, for at most two rounds. Alternative candidates are retained with the reason they were not selected. Offline preparation and online design. Offline, the system derives biological properties and records known relationships between parts. Online, it checks the request for safety, identifies the required modules, retrieves candidates, assembles and reviews designs, repairs local problems and renders the final report. The multi-agent workflow uses the same retrieval code as full_agent, so it does not change the underlying search ranking. 3.2 Integrated Evidence Library Table 2 summarizes the library. It contains 23,762 records: 6,009 biological-part records, 17,715 measured promoter–RBS combinations or literature-supported relationships, and 38 records about physical actuation methods. iGEM provides engineered parts with DNA sequences; Kosuri 2013 provides measured promoter and RBS performance; FPbase provides fluorescent proteins; and an extracted literature collection provides evidence about relationships between parts. All sources are mapped to a common evidence-record format. The 12,480 promoter–RBS records contain measured expression values from one experiment. They record the expression level observed for each pair and allow the system to compare candidate pairs on the same experimental scale. To propose a DNA construct, the system needs a DNA coding sequence rather than only a protein sequence. We used National Center for Biotechnology Information (NCBI) E-utilities [28] to add verified nucleotide coding sequences for 301 of 1,040 fluorescent proteins. The iGEM source contributes 4,000 records with DNA sequences, selected from a registry of approximately 77,000 parts. The 5,235 relationship records link a pair of biological entities to a supporting DOI and state whether the reported relationship is positive, negative or conditional. Only high- and medium-confidence records with verified DOI support are used; together they contain 865 unique DOIs. Default searches use a smaller, curated index. Large auxiliary datasets and paper-level design records are searched only when needed. This separation was introduced after adding 1,383 paper records to the default index reduced the top-five target hit rate from 15 of 21 cases (0.714) to 13 of 21 (0.619) on a retrieval regression set containing target part identifiers and hard-negative cases. Removing them from the default index restored the original score. New data are added to the default index only if this search-quality test remains unchanged. When a record lacks a verified DNA sequence, it is marked as low confidence rather than assigned an inferred sequence. Table 2: The integrated evidence library. The default search uses the curated records; large auxiliary sources are searched only when relevant. record type count DNA sequence DOI source / role curated records 748 — 737 standard parts and design records measured promoter–RBS pairs 12,480 — all Kosuri 2013 expression measurements literature relationships 5,235 — all relationships supported by 865 unique DOIs iGEM parts 4,000 all — Registry parts with DNA sequences fluorescent-protein records 1,040 301 1,020 FPbase proteins; verified DNA coding sequences where available Kosuri parts 221 all all promoters and RBSs measured on one scale actuation papers 23 — all evidence for physical actuation methods control methods 15 — — magnetic, optical or ultrasound control 3.3 Maintaining and Expanding the Library The library is versioned so that each recommended part can be traced to its source, new data do not reduce search quality, and missing categories can be measured and filled systematically. The library has three layers. The source layer preserves original identifiers, download dates and DOIs or PubMed identifiers (PMIDs). The standardized layer maps records from different sources to the same fields, such as part ID, name, function, sequence, evidence and confidence. The search layer exposes a curated default index and additional evidence that can be searched when needed. New records are cleaned and deduplicated. DNA sequences are compared in both orientations and names are matched through a synonym table. Repeated literature relationships are merged while preserving their supporting passages. Records with unclear entities or unverifiable DOIs remain available as background evidence but are not recommended as parts. Evidence is grouped into four levels: measured part combinations, verified DNA sequences, literature-supported relationships and function-based suggestions. The report states the evidence type so readers can distinguish measured results from sequence records and literature-based inference. Before release, each update is checked for valid fields, valid DNA characters, DOI/PMID format, duplicates and changes in search quality. Each version records the added and removed entries, module coverage and search metrics. Missing modules found during design, such as absent containment parts or host-specific replication origins, are recorded as priorities for the next library update. After new data are added, the system checks both search quality and coverage of the affected module. Current priorities are containment, host-specific replication origins, standardized sensor names, verified DNA coding sequences and manually checked literature relationships. 3.4 Baseline Systems To compare the complete system with simpler workflows, we use four baselines that receive the same input (Table 3). rag_only returns a retrieved list without an LLM. full_agent is a single-agent system that reranks the curated pool by keyword and intent and then produces one LLM answer; it also provides the search component used by the multi-agent system. multi_agent adds query rewriting and combines several searches. robot_design_agent adds separate handling for physical control and genetic circuits. micro_biorobot_agent sits above this ladder as the complete shared-workspace system. Table 3: Systems compared in the v1 evaluation. system workflow key mechanism rag_only none (no LLM) retrieval list + template synthesis full_agent single agent keyword-based ranking → one LLM answer multi_agent multi-pass query rewrite + combined search results robot_design two-layer physical-control routing + genetic circuit micro_biorobot shared multi-agent agents + rules + independent review + validation 3.5 Making Retrieved Parts Visible to the Model False gaps mainly arose from information loss after search. The early report generator passed only the top eight retrieved IDs to the LLM; domain-specific parts ranked lower were therefore invisible, and the model often declared the corresponding module empty. Even when an on-target part appeared in the top 8, the model still had to infer its module role. That mapping was unstable. For design, forcing a wrong part is harmful and an explicit gap is acceptable, but retrieving a usable part and then failing to name it reduces both coverage and source accuracy. A rule-based step assigns every retrieved record to one of nine biological roles (Table 4) without changing its search rank. For example, a sensor must be marked as an input and cannot also be a logic gate. The resulting part lists are supplied to the report generator for each module. If a module has at least one suitable retrieved part, the report must name it rather than describe the module as empty. Table 4: Biological roles used to organize retrieved parts. role matching information module metal sensor input part + metal cue (Hg/As/Cd/Cu) sensing metabolic sensor input part + metabolite cue (lactate/pH) sensing generic sensor input part without a logic function sensing recombinase memory recombinase type or integrase name logic/memory logic gate NOT/AND/OR/logic or CRISPR type logic/memory quorum-sensing communication quorum-sensing function or system name communication fluorescent output reporter type or fluorescent-protein record reporter kill switch toxin, kill or lysis term safety physical drive physical-control method drive control Users do not always name every required module. For example, a request may imply logic without using the phrase “logic gate”. When a required module has no assigned candidate, the system can consider a small set of broadly applicable roles, including logic, memory, generic sensing, reporting and containment (Algorithm 1). It does not apply this fallback to context-specific roles such as metal sensing, metabolic sensing, quorum sensing or physical actuation. This restriction prevented nine incorrect metabolic-part assignments observed with an unrestricted fallback. Algorithm 1 Assigning retrieved parts to required modules 1: input: retrieved IDs R, library records C, query intent q 2: coverage←coverage←\\; fallback←fallback←\\ 3: for p∈Rp∈ R do 4: r←ClassifyRole(C[p])r← ClassifyRole(C[p]) ⊳ one of nine roles, or None 5: if r=r= None then 6: continue 7: end if 8: (m,cues)←RoleToModule(r)(m,cues)← RoleToModule(r) 9: if q matches cues then ⊳ role is relevant to the request 10: append view(p)(p) to coverage[m]coverage[m] 11: else if r∈GeneralRolesr∈ GeneralRoles then ⊳ broadly applicable roles only 12: append view(p)(p) to fallback[m]fallback[m] 13: end if 14: end for 15: for module m with coverage[m]=∅coverage[m]= and fallback[m]≠∅fallback[m]≠ do 16: coverage[m]←fallback[m]coverage[m] [m] ⊳ promote only to fill an empty module 17: end for 18: within each module: prioritize role match, DOI support and search rank; keep at most 4 19: return coverage 3.6 Rule-Based Output Checks After the LLM writes the report, three rule-based checks compare its part statements with the exact set of retrieved IDs. These checks edit wording only; they do not change the retrieved candidates or their ranking. False-gap correction. If the retrieved list contains a suitable part but the report says that the module is empty, the statement is replaced with the highest-ranked suitable part. A module with no suitable retrieved part remains an explicit gap. Unsupported-part labelling. If the report presents a library part as usable even though it was not retrieved, the report marks it as an external suggestion. Exact ID matching and nearby negation checks prevent statements such as “do not use X” from being changed. Source-label correction. If the report incorrectly labels a retrieved part as external, this check changes the label to “retrieved, usable” (Algorithm 2). IDs are matched exactly, with longer IDs checked first to avoid partial matches. Algorithm 2 Correcting an external label on a retrieved part 1: input: rendered answer A, retrieved IDs R 2: ←T← forms of the “external / not retrieved” tag 3: sort R by length, descending ⊳ check longer IDs first 4: for each tag occurrence t∈t found in A at position i do 5: find the retrieved ID p∈Rp∈ R that ends exactly at i with a token boundary before it 6: if such p exists then 7: replace t with “(retrieved, usable)” 8: end if⊳ otherwise leave the external label unchanged 9: end for 10: return A Automated regression tests cover false-gap correction, external-part labelling and source-tracking correction using synthetic examples. 3.7 Building the Final Design Report Three additional steps convert retrieved parts into a readable design report. Module completion. A domain rule table identifies prerequisite modules implied by the goal but not named by the user. Required rules add a module, for example when memory implies upstream sensing or in-vivo release implies containment. Suggested rules add only a note. Report-only goals are treated as stateless and do not automatically receive a memory module. This step identifies missing modules but does not select parts. Alternative parts. Parts with the same function, evidence level and reporter colour are grouped as alternatives. If several parts can fill the same module without changing the design, the report presents them as options rather than selecting one without evidence. Reporter choices are also filtered by the requested colour. Report structure. The report begins with a design summary and then explains each module. It names candidate parts, states the evidence level, lists alternatives and describes assumptions about connections between modules. If direct compatibility evidence is unavailable, the report marks the connection as requiring validation. 4 Evaluation Setup Evaluation sets. The main comparison uses two author-developed, non-overlapping sets of 50 bio-robot design queries each. Basic Design tests module architecture and layered-circuit reasoning, including the separation of motion control from gene expression. Scenario Design asks for complete high-level designs for applications such as in-vivo tumour residence, gut living medicines, wound infection, environmental monitoring, biomanufacturing, vascular thrombosis and tissue engineering. A third, independent applied-scenario set contains different requests from the same broad application areas and is used for the blind human evaluation and component-removal study. Source-label correction is evaluated on another frozen set of 50 paired applied-scenario answers that contains the relevant source-tracking error pattern. All sets evaluate report quality rather than wet-lab performance. The questions, scoring rubrics, system outputs and aggregation code will be released with the public benchmark artifacts. Model and inference configuration. All local LLM systems use the same OpenAI-compatible vLLM service [16] backed by the Qwen3.5-27B model. The service runs on two H200 GPUs with tensor parallel size 2. Generation uses temperature 0.2; seed is not fixed, and top-p is not set explicitly and therefore follows the installed vLLM default. We retain one generation for each system–query pair. Retrieval uses BAAI/bge-m3 1024-dimensional embeddings [7] with Facebook AI Similarity Search (FAISS) [14]. A bilingual keyword and intent check makes small rule-based adjustments to the search order; no additional neural ranking model is used. Automatic scoring uses Claude Opus 4.8 [2]. The two external control groups are GPT-5.5 with the life-science skill and GPT-5.5 with the life-science skill and local RAG context: the first directly calls the GPT-5.5 model [24] with the same life-science skill, and the second adds the local RAG context to the same skill call. All local results reported here use Qwen3.5-27B. In the tables, LS means the life-science skill and LS+RAG means the same skill with local RAG context. Evaluation criteria. Each answer is scored from 0 to 10 on five criteria: completeness, whether all required modules are covered by named parts; compatibility, whether connections between modules are supported or clearly marked as uncertain; design detail, whether the report includes candidate promoters, ribosome-binding sites (RBSs), coding sequences, terminators and a replication origin; source accuracy, whether statements about retrieved parts and supporting evidence are correct; and clarity, whether a non-expert can follow the report. The overall score is assigned separately rather than calculated as the mean of these five criteria. We also report the absolute count of false-gap incidents, including source-tracking errors; because it depends on answer length and the number of checkable claims, it is not a normalized error rate. Scoring procedure. Each answer is first scored using the same instructions. A separate source check then compares every named part with the list of parts retrieved for that system. The same criteria and source check are applied to all systems. Reported values are arithmetic means across queries. The v1 study does not estimate confidence intervals or statistical significance, so differences between system means are interpreted descriptively. Paired ablation. To evaluate source-label correction, we score the same micro answers before and after the check using the same scorer and report wins, losses, ties and per-criterion Δ . Holding the answers fixed reduces variation unrelated to the output check. Regression tests. The multi-agent workflow does not modify the search implementation. Automated tests cover source tracking, module assignment and 35 output-safety cases. The retrieval regression test verifies that the stored rankings remain unchanged when orchestration and output checks are enabled. 5 Results We ask three questions: how the complete system compares with retrieval and single-pass baselines, whether a separate blind human evaluation yields a similar system-level ordering, and which components contribute most to the result. 5.1 Cross-System Comparison We evaluate seven systems on 100 author-developed evaluation queries. The collection contains two 50-question subsets. Basic Design tests module architecture and layered-circuit reasoning. Scenario Design asks for complete designs in applications including in-vivo tumour residence, gut living medicines, wound infection, environmental monitoring, biomanufacturing, vascular thrombosis and tissue engineering. Five systems come from the baseline workflows in Sec. 3.4; the other two are the external GPT-5.5 life-science controls defined above. Tables 5 and 6 report mean scores under the same evaluation criteria. Table 5: Basic Design subset (n=50n=50). Best score in each column is bold. system overall complete compat. detail source acc. clarity false gaps micro_biorobot 7.35 7.32 6.59 6.08 8.11 7.89 8 GPT-5.5+LS+RAG 6.39 6.44 5.26 4.56 8.64 8.07 0 robot_design 6.33 6.24 5.73 5.87 7.52 6.75 6 full_agent 6.11 5.82 5.08 5.11 7.80 6.91 6 GPT-5.5+LS 5.79 5.90 4.85 3.36 8.52 7.82 1 multi_agent 5.38 5.02 4.57 4.15 7.46 6.63 6 rag_only 3.66 3.83 2.15 2.62 6.48 4.55 3 Table 6: Scenario Design subset (n=50n=50). Best score in each column is bold. system overall complete compat. detail source acc. clarity false gaps micro_biorobot 8.04 8.11 7.78 7.51 8.44 8.59 2 GPT-5.5+LS+RAG 5.81 6.05 4.41 3.97 8.46 7.56 1 robot_design 5.53 5.55 5.09 4.90 7.26 6.17 15 GPT-5.5+LS 5.50 5.79 4.24 3.08 8.47 7.45 0 full_agent 5.40 5.01 4.61 4.44 7.98 6.07 5 multi_agent 4.29 3.88 3.54 2.92 7.62 5.65 3 rag_only 2.90 2.93 1.78 1.93 6.91 3.33 0 micro_biorobot_agent ranks first on both subsets. It scores 7.35 on Basic Design, 0.960.96 above the runner-up, and 8.04 on Scenario Design, 2.232.23 above the runner-up. The local-RAG-context life-science control is the strongest baseline. In this comparison, adding local RAG context is associated with a 0.30.3–0.60.6 higher overall score than the plain GPT-5.5 life-science skill. The remaining gap is concentrated in design detail and compatibility. On Scenario Design, micro scores 7.51 on design detail versus 3.97 for this baseline; on compatibility the scores are 7.78 versus 4.41. Within this benchmark, evidence retrieval alone does not produce reports with comparable overall scores. The complete system also includes module assignment, interface checking, prerequisite-module completion and explicit gap reporting. Figure 2: Mean overall scores on the two 50-query subsets. Figure 2 uses short system labels: rag is rag_only; multi is multi_agent; full is full_agent; GPT is the GPT-5.5 life-science control; robot is robot_design_agent; GPT+ctx is the GPT-5.5 control with local RAG context; and micro is micro_biorobot_agent. The suffix ctx means context. Basic and Scenario refer to the Basic Design and Scenario Design subsets. 5.2 Blind Human Evaluation We also conducted a blind human evaluation on the independent applied-scenario set (n=50n=50). The seven systems were anonymised as “Model A”–“Model G”, and the mapping was withheld until scoring was complete. The weighted rubric totals 10 points: task understanding (0–2), module decomposition (0–2), scenario fit (0–2), evidence grounding (0–1.5), feasibility (0–1.5) and safety/boundary (0–1). The retained v1 records do not state the number or background of human scorers or inter-rater statistics, so we treat this evaluation as a secondary descriptive comparison. Table 7: Mean blind human-evaluation scores on the independent applied-scenario set (n=50n=50); all seven systems were anonymised during scoring. system total task decomp. scenario evidence feasib. safety micro_biorobot 8.37 1.80 1.69 1.54 1.10 1.35 0.89 GPT-5.5+LS 7.54 1.66 1.41 1.21 1.30 1.05 0.90 GPT-5.5+LS+RAG 7.47 1.63 1.42 1.15 1.30 1.08 0.89 full_agent 7.43 1.64 1.41 1.40 0.88 1.12 0.99 robot_design 7.35 1.76 1.43 1.32 0.87 1.17 0.80 multi_agent 7.09 1.69 1.34 1.28 0.66 1.12 0.98 rag_only 5.23 1.37 1.14 0.97 0.39 0.79 0.57 micro_biorobot_agent scores 8.37, leads the next system by 0.830.83, and has the highest mean in four of six criteria. Across 50 queries, micro is the sole top-scoring system on 86% and tied for first on another 2%; 70% of its scores are at least 8, and 26% are at least 9. The two external GPT-5.5 life-science controls rank second and third, and tie for first on evidence grounding. This is consistent with their retrieval strength. We checked all 26 unique references cited by these systems against PubMed; 26/26 exist with matching titles. We therefore retain their evidence credit rather than treating it as hallucination or citation inflation. micro does not lead on literature evidence or safety. The first result reflects the external systems’ strength in literature retrieval; the second reflects limited containment parts in the current library. Across the five local systems shared with the model-scored comparison, the aggregate rankings have Spearman ρ=0.90ρ=0.90 (k=5k=5). The evaluations use different query sets and k is small, so this statistic is descriptive rank agreement rather than evidence that human and model scores are equivalent. 5.3 Correcting Source-Tracking Errors Table 8 reports the paired evaluation of source-label correction on retrieved parts. Table 8: Paired evaluation before and after the source-tracking check (n=50n=50). source acc. overall compat. complete detail clarity Δ (after −- before) +0.75 +0.32 +0.31 +0.11 +0.03 +0.02 paired overall: 32 win / 14 loss / 4 tie; target effect: 18 corrected / 32 unchanged / 0 worsened false-gap incidents: 15 → 3 (−80%-80\%) The check reduces false-gap incidents from 15 to 3, an 80% reduction. Source accuracy increases by +0.75+0.75, the largest gain; the overall score increases by +0.32+0.32, and compatibility by +0.31+0.31. The direct source-label check finds 18 corrected queries and 32 unchanged queries, with no unretrieved ID incorrectly admitted. When the complete answers are rescored, overall scores improve on 32 queries, decline on 14 and remain tied on 4. The target audit has no worsened cases; the holistic losses can reflect scoring variation outside the exact source label being corrected. 5.4 Component Ablation We use component-removal runs to examine which stages are associated with lower scores. In Table 9, Δ means “ablated −- full”. Rule-based output checks are evaluated from stored outputs, whereas module completion, review, repair and validation require fresh LLM runs. We report both mean Δ and the number of queries that receive a lower score. Table 9: Component-removal analysis on the independent applied-scenario set (n=50n=50). Δ = ablated −- matched full; negative values indicate lower observed scores after removal. ablation overall complete source acc. false gaps worse/50 −- module completion −0.91-0.91 −1.30-1.30 −0.45-0.45 7→107→10 43 −- review agents −0.78-0.78 −0.82-0.82 −0.63-0.63 3→103→10 38 −- repair step −0.65-0.65 −0.66-0.66 −0.56-0.56 4→134→13 30 −- validation checks −0.52-0.52 −0.48-0.48 −0.46-0.46 5→105→10 34 −- all three checks −0.29-0.29 −0.05-0.05 −0.63-0.63 6→116→11 28 −- unsupported-part labelling −0.25-0.25 0.000.00 −0.58-0.58 4→94→9 20 −- false-gap correction −0.05-0.05 −0.03-0.03 −0.05-0.05 6→86→8 20 Among the component-removal runs, module completion produces the largest overall decrease (−0.91-0.91) and the largest decrease in completeness (−1.30-1.30). Removing the review stage lowers source accuracy by 0.630.63 and increases false gaps from 3 to 10. The arrows show the matched-full count followed by the ablated count. These matched-full counts vary because LLM-dependent ablations were regenerated without a fixed seed; the deltas are therefore descriptive. The effects are not additive; this may reflect interaction, partial compensation or regeneration noise. The output checks have a narrower effect. Unsupported-part labelling and source-label correction mainly improve source accuracy, while false-gap correction mainly reduces false gaps. They have little effect on completeness or design detail, as expected for wording-level corrections. A representative case illustrates the role of review. Without the review agent, the system labels a logic-gate part already present in the retrieved list as “external, not retrieved” and then claims that the library has no logic-gate sequence. The full system labels the part as retrieved and usable. On this item, source accuracy falls from 10 to 5 and false gaps rise from 0 to 2 when review is removed. Regression tests also pass. The multi-agent workflow and output checks do not modify the search implementation; the source-tracking tests, 35 output-safety cases and retrieval regression tests all pass. 5.5 Example Design On a representative “multi-input, anti-interference tumour detection” query, micro first gives an overview and then names real parts module by module: a light- or hypoxia-induced sensor, a CRISPR interference (CRISPRi) logic gate, interchangeable green reporters, Kosuri RBSs, iGEM terminators and a replication origin. It also groups three reporters as interchangeable and marks a non-usable logic identifier as such. When the library contains only one real containment part, the system does not invent a safety module. It marks the safety module as a genuine gap, preserving source accuracy and making the limitation clear to the reader. 6 Limitations and Discussion This study evaluates high-level design reports rather than functional biological constructs. We do not compile the generated designs with Cello, encode them in SBOL, simulate their dynamics, or validate them in cell-free or wet-lab systems. The reported metrics therefore measure report quality, and all designs require expert review before experimental use. Library coverage is uneven. Sensing (69), logic (71) and output (47, plus 1,040 fluorescent-protein coding-sequence entries) have substantially more curated candidates than safety/containment (1) and host/replication-origin (7). These gaps limit the completeness of some designs and motivate targeted expansion with evidence-backed kill switches and host-specific backbone parts. The benchmark was developed by the authors, and each system–query pair was generated once without a fixed seed. The study therefore does not estimate run-to-run variance, confidence intervals or the statistical significance of mean differences. Evaluation also relies mainly on a model-based scorer; the blind human evaluation provides an additional ranking comparison, but it uses a different query set and the retained records do not support inter-rater analysis. In addition, system budgets are not fully matched, so the results compare complete systems rather than isolating the effect of agent count alone. External expert benchmarks, repeated runs and budget-matched comparisons are important directions for future work. All local evaluations in this paper use the Qwen3.5-27B v1 system. The system is restricted to high-level module choices, evidence and uncertainty and does not provide executable experimental procedures. It is not intended for direct clinical use, environmental release or autonomous experimentation. 7 Conclusion We present micro_biorobot_agent, an offline multi-agent system that converts bio-robot requirements into modular design reports with traceable evidence and explicit uncertainty. The system combines a shared structured design record, specialized agents, deterministic validators and rule-based output checks. It has the highest observed mean overall score among the seven systems evaluated on both 50-query subsets and scores 8.04 on Scenario Design, 2.23 points above the runner-up. Source-label correction reduces false-gap incidents by 80% and increases source accuracy by +0.75+0.75. These results document the observed performance of the Qwen3.5-based v1 system and motivate broader benchmarking and experimental validation in future work. References [1] Y. Alapan, O. Yasa, B. Yigit, I. C. Yasa, P. Erkoc, and M. Sitti (2019) Microrobotics and microorganisms: biohybrid autonomous cellular robots. Annual Review of Control, Robotics, and Autonomous Systems 2, p. 205–230. External Links: Document Cited by: §1. [2] Anthropic (2026) Claude opus 4.8 system card. External Links: Link Cited by: §4. [3] D. A. Boiko, R. MacKnight, B. Kline, and G. Gomes (2023) Autonomous chemical research with large language models. Nature 624 (7992), p. 570–578. External Links: Document Cited by: §2. [4] A. M. Bran, S. Cox, O. Schilter, C. Baldassari, A. D. White, and P. Schwaller (2024) Augmenting large language models with chemistry tools. Nature Machine Intelligence 6, p. 525–535. External Links: Document Cited by: §2. [5] B. Canton, A. Labno, and D. Endy (2008) Refinement and standardization of synthetic biological parts and devices. Nature Biotechnology 26 (7), p. 787–793. External Links: Document Cited by: §1, §2. [6] C. T. Y. Chan, J. W. Lee, D. E. Cameron, C. J. Bashor, and J. J. Collins (2016) “Deadman” and “Passcode” microbial kill switches for bacterial containment. Nature Chemical Biology 12 (2), p. 82–86. External Links: Document Cited by: §1. [7] J. Chen, S. Xiao, P. Zhang, K. Luo, D. Lian, and Z. Liu (2024) M3-embedding: multi-linguality, multi-functionality, multi-granularity text embeddings through self-knowledge distillation. In Findings of the Association for Computational Linguistics: ACL 2024, p. 2318–2335. External Links: Document Cited by: §4. [8] M. O. Din, T. Danino, A. Prindle, M. Skalak, J. Selimkhanov, K. Allen, E. Julio, E. Atolia, L. S. Tsimring, S. N. Bhatia, and J. Hasty (2016) Synchronized cycles of bacterial lysis for in vivo delivery. Nature 536 (7614), p. 81–85. External Links: Document Cited by: §1. [9] D. Endy (2005) Foundations for engineering biology. Nature 438 (7067), p. 449–453. External Links: Document Cited by: §1. [10] L. D. Erman, F. Hayes-Roth, V. R. Lesser, and D. R. Reddy (1980) The HEARSAY-I speech-understanding system: integrating knowledge to resolve uncertainty. ACM Computing Surveys 12 (2), p. 213–253. External Links: Document Cited by: §2. [11] M. Galdzicki, K. P. Clancy, E. Oberortner, M. Pocock, J. Y. Quinn, C. A. Rodriguez, N. Roehner, M. L. Wilson, L. Adam, J. C. Anderson, B. A. Bartley, J. Beal, D. Chandran, J. Chen, D. Densmore, D. Endy, R. Grünberg, J. Hallinan, N. J. Hillson, J. D. Johnson, A. Kuchinsky, M. Lux, G. Misirli, J. Peccoud, H. A. Plahar, E. Sirin, G. V. Stan, A. Villalobos, A. Wipat, J. H. Gennari, C. J. Myers, and H. M. Sauro (2014) The synthetic biology open language (SBOL) provides a community standard for communicating designs in synthetic biology. Nature Biotechnology 32 (6), p. 545–550. External Links: Document Cited by: §1, §2. [12] iGEM Foundation (2026) Registry of standard biological parts. Note: https://registry.igem.org/Accessed 18 June 2026 Cited by: §1, §2. [13] Z. Ji, N. Lee, R. Frieske, T. Yu, D. Su, Y. Xu, E. Ishii, Y. Bang, D. Chen, W. Dai, A. Madotto, and P. Fung (2023) Survey of hallucination in natural language generation. ACM Computing Surveys 55 (12), p. 1–38. External Links: Document Cited by: §2. [14] J. Johnson, M. Douze, and H. Jégou (2021) Billion-scale similarity search with GPUs. IEEE Transactions on Big Data 7 (3), p. 535–547. External Links: Document Cited by: §2, §4. [15] S. Kosuri, D. B. Goodman, G. Cambray, V. K. Mutalik, Y. Gao, A. P. Arkin, D. Endy, and G. M. Church (2013) Composability of regulatory sequences controlling transcription and translation in Escherichia coli. Proceedings of the National Academy of Sciences of the United States of America 110 (34), p. 14024–14029. External Links: Document Cited by: §2. [16] W. Kwon, Z. Li, S. Zhuang, Y. Sheng, L. Zheng, C. H. Yu, J. E. Gonzalez, H. Zhang, and I. Stoica (2023) Efficient memory management for large language model serving with PagedAttention. In Proceedings of the 29th Symposium on Operating Systems Principles, p. 611–626. External Links: Document Cited by: §4. [17] T. J. Lambert (2019) FPbase: a community-editable fluorescent protein database. Nature Methods 16 (4), p. 277–278. External Links: Document Cited by: §1, §2. [18] J. M. Laurent, J. D. Janizek, M. Ruzo, M. M. Hinks, M. J. Hammerling, S. Narayanan, M. Ponnapati, A. D. White, and S. G. Rodriques (2024) LAB-Bench: measuring capabilities of language models for biology research. arXiv preprint arXiv:2407.10362. External Links: Document Cited by: §2. [19] P. Lewis, E. Perez, A. Piktus, F. Petroni, V. Karpukhin, N. Goyal, H. Küttler, M. Lewis, W. Yih, T. Rocktäschel, S. Riedel, and D. Kiela (2020) Retrieval-augmented generation for knowledge-intensive NLP tasks. In Advances in Neural Information Processing Systems, Vol. 33, p. 9459–9474. External Links: Link Cited by: §2. [20] J. Maynez, S. Narayan, B. Bohnet, and R. McDonald (2020) On faithfulness and factuality in abstractive summarization. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, p. 1906–1919. External Links: Document Cited by: §2. [21] J. A. McLaughlin, C. J. Myers, Z. Zundel, G. Misirli, M. Zhang, I. D. Ofiteru, A. Goni-Moreno, and A. Wipat (2018) SynBioHub: a standards-enabled design repository for synthetic biology. ACS Synthetic Biology 7 (2), p. 682–688. External Links: Document Cited by: §1, §2. [22] V. K. Mutalik, J. C. Guimaraes, G. Cambray, C. Lam, M. J. Christoffersen, Q. Mai, A. B. Tran, M. Paull, J. D. Keasling, A. P. Arkin, and D. Endy (2013) Precise and reliable gene expression via standard transcription and translation initiation elements. Nature Methods 10 (4), p. 354–360. External Links: Document Cited by: §2. [23] A. A. K. Nielsen, B. S. Der, J. Shin, P. Vaidyanathan, V. Paralanov, E. A. Strychalski, D. Ross, D. Densmore, and C. A. Voigt (2016) Genetic circuit design automation. Science 352 (6281), p. aac7341. External Links: Document Cited by: §2. [24] OpenAI (2026) Introducing GPT-5.5. External Links: Link Cited by: §4. [25] Y. Qu, K. Huang, M. Yin, K. Zhan, D. Liu, D. Yin, H. C. Cousins, W. A. Johnson, X. Wang, M. Shah, R. B. Altman, D. Zhou, M. Wang, L. Cong, et al. (2026) CRISPR-GPT for agentic automation of gene-editing experiments. Nature Biomedical Engineering 10, p. 245–258. External Links: Document Cited by: §2. [26] Qwen Team (2026) Qwen3.5: towards native multimodal agents. External Links: Link Cited by: §1. [27] S. Robertson and H. Zaragoza (2009) The probabilistic relevance framework: BM25 and beyond. Foundations and Trends in Information Retrieval 3 (4), p. 333–389. External Links: Document Cited by: §2. [28] E. Sayers (2022) A general introduction to the E-utilities. Note: NCBI BookshelfUpdated 17 November 2022 External Links: Link Cited by: §3.2. [29] T. Schick, J. Dwivedi-Yu, R. Dessì, R. Raileanu, M. Lomeli, E. Hambro, L. Zettlemoyer, N. Cancedda, and T. Scialom (2023) Toolformer: language models can teach themselves to use tools. In Advances in Neural Information Processing Systems, Vol. 36, p. 68539–68551. External Links: Document Cited by: §2. [30] K. Swanson, W. Wu, N. L. Bulaong, J. E. Pak, and J. Zou (2025) The virtual lab of AI agents designs new SARS-CoV-2 nanobodies. Nature 646, p. 716–723. External Links: Document Cited by: §2. [31] Z. Wang, Q. Jin, C. Wei, S. Tian, P. Lai, Q. Zhu, C. Day, C. Ross, R. Leaman, and Z. Lu (2025) GeneAgent: self-verification language agent for gene-set analysis using domain databases. Nature Methods 22, p. 1677–1685. External Links: Document Cited by: §2, §2. [32] Q. Wu, G. Bansal, J. Zhang, Y. Wu, B. Li, E. Zhu, L. Jiang, X. Zhang, S. Zhang, J. Liu, A. H. Awadallah, R. W. White, D. Burger, and C. Wang (2024) AutoGen: enabling next-gen LLM applications via multi-agent conversations. In First Conference on Language Modeling, External Links: Link Cited by: §2. [33] S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y. Cao (2023) ReAct: synergizing reasoning and acting in language models. In International Conference on Learning Representations, External Links: Link Cited by: §2. [34] L. Zheng, W. Chiang, Y. Sheng, S. Zhuang, Z. Wu, Y. Zhuang, Z. Lin, Z. Li, D. Li, E. P. Xing, H. Zhang, J. E. Gonzalez, and I. Stoica (2023) Judging LLM-as-a-judge with MT-Bench and chatbot arena. In Advances in Neural Information Processing Systems, Vol. 36, p. 46595–46623. External Links: Document Cited by: §2. Appendix A Reproducibility and Artifact Scope The frozen v1 records include the evaluation questions, library version, model configuration, system outputs, scoring items and aggregation code used for the reported tables. The questions, rubrics, outputs and aggregation code will be released as a public benchmark package. Automated tests cover assignment of retrieved parts to modules, false-gap correction, source-label correction, library separation, module completion and 35 output-safety cases. Paired comparisons reuse the same answers and retrieved part sets. Because generation did not use a fixed seed, the frozen outputs define the exact results reported in this paper. The public package will not redistribute third-party paper PDFs. Appendix B Author Contributions and AI Use Disclosure Author contributions. Yujun Chen, Tianle Li, Jiayu Chen and Zhen Yin contributed to the study design, system development, data analysis and manuscript preparation. All authors reviewed and approved the final manuscript. AI use disclosure. AI tools assisted with data cleaning, code drafting, output scoring and language editing. The authors defined the research questions, designed the system and evaluation, checked the data and results, verified the references and take responsibility for the manuscript.