How Co-Scientist orchestrates multi-agent hypothesis generation in biomedical research

Co-Scientist is a multi-agent AI system built on Gemini. A researcher provides a research goal in natural language. A Supervisor agent parses the goal into a research plan, then orchestrates a coalition of specialised agents: Generation (explores literature and produces initial hypotheses through simulated scientific debates), Reflection (acts as a peer reviewer examining correctness, quality, and safety), Ranking (evaluates proposals using Elo-based tournament scoring), Evolution (refines and combines top-ranked hypotheses), and Meta-review (synthesises insights from the debate cycle and produces the final research proposal).

The system uses an asynchronous task execution framework that scales test-time compute - allocating more processing time to harder problems, similar to the approach used in OpenAI's o1. Unlike linear LLM prompting, Co-Scientist's freeform planner breaks down high-level research goals into executable steps, coordinating agents to run in parallel and explore multiple avenues simultaneously. Generated ideas are iteratively refined, critiqued, and evolved into new hypotheses.

The validation studies are concrete. In drug repurposing for liver fibrosis, Co-Scientist recommended targeting epigenomic modifiers. Two of three recommended drugs exhibited significant anti-fibrotic activity in human hepatic organoids; one - Vorinostat, an FDA-approved anti-cancer treatment - reduced TGFβ-induced chromatin structural changes by 91% (Guan et al., Advanced Science, 2025; DOI: 10.1002/advs.202508751). In antimicrobial resistance research, Co-Scientist independently proposed the same mechanism of cross-species gene transfer via chimeric phage-inducible chromosomal islands that researchers at Imperial College London had spent years developing experimentally (Penadés et al., Cell, 2025; DOI: 10.1016/j.cell.2025.08.018).

How BioSkepsis retrieves and reasons over biomedical literature for researchers

BioSkepsis retrieval is weighted by Gene Ontology terms, MeSH descriptors, gene symbols, and pathway relationships. A query about AMPK activation and its downstream effects on hepatic lipogenesis returns papers linked by the biological concepts involved, not just papers whose abstracts contain matching keywords. This biology-native retrieval layer is what separates BioSkepsis from both general-purpose LLMs and multi-agent hypothesis systems that treat biomedical papers as undifferentiated text for idea mining.

Answers are grounded in full text including methods, controls, and supplementary data. Every claim links back to the exact passage in the retrieved paper. When evidence is insufficient, BioSkepsis declines to answer rather than generating a plausible-sounding response. The research landscape graph classifies papers by structural role (Foundational, Hub, Bridge, and Novel) and draws on Semantic Scholar's 214M+ corpus for landscape expansion.

BioSkepsis also performs citation verification: a seven-step pipeline (Steps A through G) that audits whether cited papers in an existing document actually support the claims they are attached to. This is a forensic capability; it checks the integrity of literature use, not just its availability.

The EFEVRE ecosystem: physical execution, documentation, and interpretation in one loop

BioSkepsis is one of three products built by EFEVRE TECH LTD. The other two address problems that Co-Scientist does not attempt to solve.

AMGEL (AutoMated GEneral Laboratory) is a patent-pending robotic platform that automates all mainstream wet-bench procedures on one device: pipetting, centrifugation, heating/cooling, magnetic bead isolation, cold storage, and PCR. It runs 860+ pre-set protocols with 24/7 unattended operation. AMGEL physically executes the experiments that tools like Co-Scientist can only propose computationally.

VITALE (Versatile Integrated Technology Advancing Life-Science Exploration) is the software layer that records every step, parameter, timestamp, and result during AMGEL operation. Its Protocol Designer, Personal Labbook, Scheduling Calendar, and Protocol Library with community scoring ensure that no experimental detail goes undocumented. VITALE addresses the documentation pillar of the reproducibility crisis: even when experiments are performed correctly, poor reporting in publications makes replication structurally impossible.

Together, the three products form a closed loop: AMGEL removes human error from execution, VITALE removes human error from documentation, and BioSkepsis removes human bias from interpretation. Co-Scientist's architecture does not include either a physical execution layer or a documentation integrity layer.

Feature comparison: BioSkepsis (EFEVRE ecosystem) vs Co-Scientist (Google DeepMind)

Where Co-Scientist leads: autonomous hypothesis generation at scale in biomedical research

Co-Scientist's primary advantage is the depth and breadth of its hypothesis exploration. The multi-agent debate architecture - where Generation, Reflection, Ranking, and Evolution agents run in parallel, with an adaptive Supervisor orchestrating the workflow - allows the system to explore thousands of research directions simultaneously. This is not prompt engineering; it is a structured search over hypothesis space, with each candidate hypothesis subjected to automated peer review and iterative refinement.

The Elo-based tournament ranking is a meaningful design choice. Rather than presenting hypotheses in the order they were generated, Co-Scientist forces candidates to compete head-to-head, with the Ranking agent evaluating novelty, feasibility, and scientific rigour. The surviving hypotheses have been stress-tested through multiple rounds of critique before reaching the researcher.

The validation results speak for themselves. In the liver fibrosis study, Co-Scientist recommended epigenomic modifier targets that two independent human expert panels had not prioritised; two of three recommended drugs showed significant anti-fibrotic activity. In the antimicrobial resistance study, Co-Scientist independently arrived at the same mechanism - chimeric phage-inducible chromosomal islands enabling cross-species gene transfer - that Imperial College researchers had developed through years of experimental work. General-purpose LLMs from OpenAI, Anthropic, DeepSeek, and Google's own Gemini 2.0 model did not produce these hypotheses when tested on the same prompts.

The integration with Google's broader AI-for-science ecosystem adds infrastructure that no standalone platform can match. AlphaFold for protein structure prediction, AlphaGenome for cancer mutation analysis, and ERA for empirical research assistance are all part of the same Gemini for Science suite. For teams already in the Google Cloud ecosystem, this integration is a significant advantage.

Where BioSkepsis leads: citation integrity, reproducibility, and structured evidence reasoning in the life sciences

BioSkepsis occupies territory Co-Scientist does not enter. Citation verification, the seven-step audit pipeline that checks whether cited papers actually support the claims they are attached to, has no equivalent in the Co-Scientist agent stack. Co-Scientist's agents synthesise literature to generate new hypotheses; they do not audit the integrity of citation use in existing publications. For researchers, reviewers, and editors concerned with the accuracy of the scientific record, this is a meaningful gap.

The research landscape graph, with its classification of papers into Foundational, Hub, Bridge, and Novel roles via citation network analysis, provides structural understanding of a field that Co-Scientist's agents do not surface. A researcher mapping the state of a sub-discipline - identifying which seminal papers anchor the field and which bridge papers connect separate domains - needs this kind of analysis. Co-Scientist reads papers to extract hypotheses; BioSkepsis maps the relationships between papers to reveal field structure.

Biology-native retrieval is the most architecturally consequential difference. BioSkepsis indexes papers by Gene Ontology terms, MeSH descriptors, gene symbols, and pathway relationships. A query about AMPK activation via LKB1 returns different papers than a query about AMPK activation via CaMKK2, because the retrieval layer understands the biological distinction. Co-Scientist's Generation agent uses web search - broad and flexible, but not structured by biological ontology. For queries where biological specificity determines relevance, domain-specific retrieval outperforms keyword and embedding approaches.

The EFEVRE ecosystem's physical layer is the most decisive differentiator. AMGEL runs 860+ laboratory protocols autonomously, 24 hours a day. VITALE records every parameter. No AI-only platform, including Co-Scientist, addresses the reproducibility crisis at the bench. Co-Scientist proposes experiments; the EFEVRE ecosystem executes, documents, and interprets them.

Where BioSkepsis and Co-Scientist overlap, and how they complement each other in biomedical research

The overlap is in literature synthesis and hypothesis generation. Both BioSkepsis and Co-Scientist's Generation agent read biomedical papers and produce outputs grounded in published evidence. A researcher querying either system about a signalling pathway will get a synthesised response with references. On this axis, the two tools are comparable in intent if not in architecture.

The architectural difference matters. Co-Scientist uses multi-agent debate and Elo ranking to push hypotheses toward novelty - the system is designed to produce ideas that go beyond what the literature explicitly states. BioSkepsis uses biology-native retrieval and full-text reasoning to produce syntheses that stay strictly within the evidence - every claim traceable to a specific passage and PMID. One system is optimised for exploration; the other for verification.

The difference in retrieval matters. BioSkepsis uses a biology-native knowledge graph weighted by Gene Ontology, MeSH, and gene symbols. Co-Scientist's agents use LLM-driven web search. For queries where biological specificity determines relevance (e.g., distinguishing papers about TGFβ signalling via SMAD2/3 from those about TGFβ signalling via non-canonical pathways), domain-specific retrieval outperforms general search.

The complementary workflow is clear. Co-Scientist generates a novel hypothesis about an epigenomic target for liver fibrosis. BioSkepsis verifies the underlying evidence - checking whether the cited papers actually support the mechanistic claims, mapping the citation network to identify foundational studies and knowledge gaps, and producing an auditable synthesis. AMGEL executes the proposed assays with full parameter control. VITALE records every experimental step. Each system handles a distinct phase; none duplicates another's core job.

Who should use which tool in biomedical and life-science research

Frequently asked questions