A definition that excludes a chatbot
Gao and colleagues (Cell 2024) describe the target as AI scientists: systems capable of skeptical learning and reasoning that empower biomedical research through collaborative agents integrating AI models and biomedical tools with experimental platforms. The emphasis in that framing is on integration and skepticism - agents that use instruments and databases, and that argue with each other's conclusions.
The authors position these agents as combining human creativity with machine analysis across applications from virtual cell simulation to therapeutic development, and explicitly as tools for knowledge integration and continuous learning rather than replacements for human researchers. That last clause is the part most product descriptions drop.
A single model answering questions in a chat window is not an agentic pipeline, however capable the model. The pipeline begins when there is more than one role, a mechanism for one role to reject another's output, and a persistent record of what was decided.
What has actually been demonstrated
The clearest published demonstration is Coscientist (Boiko et al., Nature 2023), a GPT-4-driven system that autonomously designs, plans and performs complex experiments by combining a language model with internet and documentation search, code execution and experimental automation. It was evaluated across six diverse tasks, including the successful reaction optimisation of palladium-catalysed cross-couplings.
That is a chemistry result, not a drug-discovery result, and the distinction is worth preserving. The tasks had fast, unambiguous readouts from automated equipment. Biomedical questions usually have slow, contested readouts measured in animals, patients and years, which is exactly the regime where an autonomous loop has no corrective signal.
So the honest summary of the published state of the art is: agentic systems can run a closed experimental loop where the loop closes quickly, and can do broad structured evidence work everywhere else. Claims beyond that are projections.
The anatomy of a working pipeline
Most functioning designs converge on the same components, whatever they are called.
- An ingestion layer that pulls from named public sources and records what it read and when
- Specialist agents with narrow scopes - a literature agent, a structure agent, a pharmacology agent - each with stated inputs, outputs and limitations
- An orchestrator that routes work, resolves conflicts between specialists and holds the overall plan
- An audit or critic layer that reviews outputs for causal, mechanistic, temporal and consistency failures before anything is released
- Named human decision rights: which decisions, such as scope, kill criteria and regulatory commitments, an agent may never make
- An immutable record linking every claim to its evidence, its evidence level and the version of the agent that produced it
Where these pipelines fail
Speed is the hazard. An agentic system can produce in an afternoon more analysis than a team can check in a week, and the failure modes are not the obvious ones. Confident fabrication of a citation is easy to catch. A subtly leaked label, an analysis chosen after seeing the result, or an analogy transferred across species without a flag are not.
Kapoor and Narayanan (Patterns 2023) documented leakage across 17 fields and 294 papers, with eight distinct types - and showed in their own reproduction that correcting the errors erased the apparent advantage of complex models. An agentic pipeline that generates its own features and its own evaluation splits can reproduce every one of those eight types faster than a human team could.
Error compounding is the second structural risk. When agent B treats agent A's output as an input rather than a hypothesis, a soft error becomes a hard premise. This is the argument for a critic layer that sees the chain rather than the conclusion, and for a human gate at each escalation of evidence.
What the pipeline is for
Used well, the value is coverage, not autonomy. Agents are good at surveying a large candidate space, reading more of the literature than a team can, maintaining a consistent evidence standard across hundreds of claims, and producing a ranked shortlist with its reasoning attached. People are good at deciding what is worth doing, what a surprising result means, and when to stop.
If the output of a biomedical question will support a regulatory submission, the credibility of the pipeline is assessed for one narrow context of use at a time, in line with the Food and Drug Administration's January 2025 draft guidance on artificial intelligence in regulatory decision-making. An agentic system does not get a general clearance. Each job it does is credible, or is not, on its own evidence.
Atlas Bio runs a multi-tier agentic stack - scientific-direction agents, an orchestrator, specialist agents, an audit layer and a human-in-the-loop board - in which a named set of decisions including scope, kill-criteria enforcement, translational gates and regulatory commitment is non-delegable to agents, and every advance between evidence tiers requires a signed human review.
How is an agentic pipeline different from a model with retrieval?
Retrieval adds evidence to a single answer. An agentic pipeline adds roles: separate agents that plan, execute, audit and escalate, with the authority to block a release. The audit layer and the decision rights, not the retrieval, are what make it agentic.
Can agents run experiments without supervision?
A published system has autonomously designed and executed chemistry experiments with automated readouts. Biomedical work with slow, contested endpoints has no comparable demonstration, and the human decisions about scope, kill criteria and clinical commitment remain non-delegable.
What is the main risk of running more agents?
Volume outpacing verification. More agents produce more claims, and the subtle failures - leaked labels, post hoc analysis choices, analogies carried across species - scale with output. A critic layer and human gates have to scale with the agent count, or the pipeline just produces mistakes faster.
- Gao S et al. Empowering biomedical discovery with AI agents. Cell 2024 (PMID 39486399)
- Boiko DA, MacKnight R, Kline B, Gomes G. Autonomous chemical research with large language models. Nature 2023 (PMC10733136)
- Kapoor S, Narayanan A. Leakage and the reproducibility crisis in machine-learning-based science. Patterns 2023 (PMC10499856)
- FDA. Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products. Draft guidance, January 2025
Talk to Atlas Bio.
Sponsors, researchers and clinicians: tell us the decision you are trying to make and we will show what the platform can and cannot answer.
Contact Us →