An Evidence-Grounded Multi-Agent System for High-Level Bio-Robot Design

Yujun Chen; Tianle Li; Jiayu Chen; Zhen Yin

cs.MAAug 20, 2026

An Evidence-Grounded Multi-Agent System for High-Level Bio-Robot Design

Authors: 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, 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.

Figures & tables

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.

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

Table 1: Main stages of the multi-agent workflow.

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

Table 2: The integrated evidence library. The default search uses the curated records; large auxiliary sources are searched only when relevant.

system	workflow	key mechanism
rag_only	none (no LLM)	retrieval list + template synthesis
full_agent	single agent	keyword-based ranking $\rightarrow$ 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

Table 3: Systems compared in the v1 evaluation.

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

Table 4: Biological roles used to organize retrieved parts.

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

Table 5: Basic Design subset (

n=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

Table 6: Scenario Design subset (

n=50

). Best score in each column is bold.

Figure 2: Mean overall scores on the two 50-query subsets.

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

Table 7: Mean blind human-evaluation scores on the independent applied-scenario set (

n=50

); all seven systems were anonymised during scoring.

	source acc.	overall	compat.	complete	detail	clarity
$\Delta$ (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 $\rightarrow$ 3 ( $-80\%$ )

Table 8: Paired evaluation before and after the source-tracking check (

n=50

ablation	overall	complete	source acc.	false gaps	worse/50
$-$ module completion	$-0.91$	$\mathbf{-1.30}$	$-0.45$	$7{\to}10$	43
$-$ review agents	$-0.78$	$-0.82$	$-0.63$	$3{\to}10$	38
$-$ repair step	$-0.65$	$-0.66$	$-0.56$	$4{\to}13$	30
$-$ validation checks	$-0.52$	$-0.48$	$-0.46$	$5{\to}10$	34
$-$ all three checks	$-0.29$	$-0.05$	$\mathbf{-0.63}$	$6{\to}11$	28
$-$ unsupported-part labelling	$-0.25$	$0.00$	$\mathbf{-0.58}$	$4{\to}9$	20

Table 9: Component-removal analysis on the independent applied-scenario set (

n=50

\Delta

= ablated

-

matched full; negative values indicate lower observed scores after removal.

Robotics

Robotics studies the design and control of machines that perceive, move, and act in physical environments.

See all 751 papers

May 7, 2026

BioResearcher: Scenario-Guided Multi-Agent for Translational Medicine

Translational medicine turns underspecified development goals into evidence synthesis that must combine literature, trials, patents, and quantitative multi-omics analysis while preserving identifiers, uncertainty, and retrievable provenance. General-purpose foundation models and off-the-shelf tool-augmented or multi-agent systems are not built for this: they tend to produce single-shot answers or run open-endedly, and fall short on the auditable, scenario-specific workflows that heterogeneous biomedical sources demand. This paper introduces Ingenix BioResearcher, a scenario-guided multi-agent system that maps queries to versioned research playbooks, delegates to specialized subagents over 30+ tools and machine-learning endpoints, mixes structured database access with sandboxed code for genome-scale analyses, and applies claim-level multi-model reconciliation before editorial assembly. We evaluate BioResearcher across unit-level capabilities, open-ended biomedical reasoning, and end-to-end clinical discovery. It leads evaluated baselines on 109 single-step tests (83.49% pass rate; 0.892 average score), achieves strong biomedical benchmark performance (89.33% on BixBench-Verified-50 and the top 0.758 mean score on BaisBench Scientific Discovery), and leads on a 30-query clinical end-to-end benchmark with the highest positive hit rate (74.7%

\pm

3.3%) and negative clear rate (96.8%

\pm

0.2%). These results show broad, competitive performance across unit-level, open-ended, and end-to-end clinical evaluations.

Remigiusz Kinas, Joanna Krawczyk, Rafał Powalski +6

Ingenix.AI, Warsaw, Poland

Multi-Agent Orchestration AI Agents for Scientific Discovery Healthcare Scientific Workflow Automation

Ask PDF

May 13, 2026

Do Biological Structural Guarantees Earn Their Complexity?

Biologically-inspired AI agent frameworks claim reliability benefits through structural guarantees adapted from gene regulatory networks, immune systems, and metabolic control. These claims are rarely tested empirically against simpler alternatives. We present three deep benchmarks: metabolic priority gating, autoinducer-based quorum sensing, and Bayesian stagnation detection, each comparing a biologically-grounded implementation against a naive non-biological alternative and an ablated control, across 1,000 trials per seed and 10 seeds (10M+ data points total).

Bogdan Banu

AI Agent Reliability AI Agent Benchmarks

Ask PDF

Jun 19, 2026

BioInsight: Multi-Agent Orchestration for Interactive Biomedical Knowledge Discovery

Biomedical researchers increasingly use AI-generated analyses and reports to interpret protein-level signals, but static outputs are often insufficient for research decision-making, where users need to inspect evidence, assess uncertainty, compare mechanisms, and refine hypotheses. We present BioInsight, a multi-agent system that moves from static biomedical report generation to interactive evidence-centered interactive interface generation. Given a disease name, a protein association table, and optional cohort metadata, BioInsight organizes disease-specific evidence through typed intermediate artifacts, including ranked pathways, literature evidence packets, protein-level reasoning notes, citation-grounded reports, dashboard schemas, and rendered interactive interfaces. The system decomposes evidence retrieval from mechanistic reasoning, normalizes citations through deterministic components, and converts the same structured evidence used in the report into an interactive interface. We evaluate BioInsight on standardized biomedical QA, challenging protein-function reasoning, and end-to-end biomedical evidence synthesis. Results show that BioInsight achieves best, and suggest that biomedical AI systems should move beyond text-only and static reports toward provenance-preserving, interactive evidence artifacts.

Jieyi Wang, Bingxuan Li, Nanyi Jiang +9

1Peking University · University of Illinois at Urbana-Champaign · University of Pennsylvania +2

AI for Science Multi-Agent Orchestration Biomedical QA Evidence-Grounded Generation

Ask PDF

Abstract

Robotics

Explore similar work

BioResearcher: Scenario-Guided Multi-Agent for Translational Medicine

Do Biological Structural Guarantees Earn Their Complexity?

BioInsight: Multi-Agent Orchestration for Interactive Biomedical Knowledge Discovery