Agentic AI in Drug Discovery: What Agents Actually Do

The lab meeting starts in an hour. Your team has a table of compounds, results from several experiments, and a folder of papers. You want to decide which candidates deserve more work. First, someone has to check the records, compare the evidence, and explain what is still missing.

An AI assistant can help discuss that task. An agent can also work through parts of it: open the available records, run an analysis, notice a problem, and change its next step. That ability to connect actions is what makes agentic AI interesting for research.

The useful question is what work the system can complete, what evidence supports its output, and where a scientist needs to decide. This guide follows a small research task to make those differences concrete.

Chatbots, workflows, and agents

A chatbot is a conversational interface. A workflow is a sequence of steps defined in advance. An AI agent is a system that can choose actions toward a goal, use the tools available to it, and respond to the results.

These descriptions can overlap. A chat window can be the front door to an agent. An agent can call a fixed workflow to perform a reliable calculation. The interface alone does not tell you how much work happens behind it.

A fixed workflow can be the right choice when a task is well understood and should run the same way every time. An agent becomes useful when the next step depends on what the previous one revealed. If an assay record lacks units, for example, the agent should resolve that gap before comparing it with other measurements.

Where reasoning fits

A model can work through a difficult question without taking action outside its answer. Giving it tools adds another capability: it can obtain information or perform a calculation, then use that result to decide what comes next.

The ReAct research paper explored this pattern by combining reasoning with actions and feedback from an external environment. It is a useful foundation for understanding agents. Its experiments were not a demonstration of drug discovery, so they should not be read as evidence that an agent can run a medicine-development program.

Better reasoning can help a system choose a useful action. The quality of that action still depends on the data, the tool, and the way the result is checked.

The work loop behind an AI agent

An agent needs a goal, access to relevant information, tools it is permitted to use, and a way to tell when it should stop. It also needs a working record of what has already happened. Otherwise, it may repeat a failed step or make a recommendation that ignores an earlier result.

The important part of this loop is the check after an action. A tool call returning successfully means that software ran. It does not establish that the input was appropriate or that the scientific conclusion is sound.

For example, a property model might accept a molecular structure and return a prediction. The team still needs to know whether that model is suitable for the kind of molecule being studied and whether the prediction is useful for the decision at hand.

A worked example: preparing a compound review

Suppose a scientist asks: “Review these 24 compounds and propose up to three for our next discussion. Use the available activity data, flag missing evidence, and explain the trade-offs.”

This is an invented teaching example. The numbers and steps illustrate a possible workflow, rather than results from an AgentCures program.

Start with an explicit task

The scientist defines what matters: the intended target, the properties of interest, which records may be used, and what the agent should deliver. “Up to three” is deliberate. The agent should be able to return fewer candidates if the evidence is too weak.

The stopping point is also clear: prepare a proposal for review. Ordering compounds or sending instructions to a laboratory would be separate actions with their own permissions.

Check that the records can be compared

The agent checks compound identifiers, molecular structures, assay descriptions, and units. An assay is a test used to measure a biological or chemical effect. Two columns labeled “activity” may describe different tests, so putting them in the same ranking could hide an important difference.

If a record is incomplete, the agent can look for the original source or flag the gap. A plausible guess is not an acceptable substitute for a missing measurement.

Use tools for specific questions

A chemistry tool might calculate a molecular descriptor. A predictive model might estimate a property that has not been measured. A literature search might find a relevant experiment.

These outputs should remain distinguishable. A calculation from a structure, a model prediction, and a laboratory measurement have different meanings. The agent’s job includes carrying those distinctions into its report, along with the inputs and tool versions needed to check the work.

Draft a recommendation with its limits visible

The final proposal explains why each candidate deserves attention, which evidence supports that view, and what would change the recommendation. It also records exclusions. If a compound was left out because its identity could not be confirmed, that should be visible to the scientist.

An honest report might say that two candidates are ready for discussion and a third needs its assay record checked. Filling every requested slot would make the report look complete while leaving the underlying question unresolved.

What should happen when the work goes wrong?

Research rarely follows a tidy script. Missing data, unavailable tools, and contradictory results are useful tests of an agent’s behavior. Open the scenarios below to explore how the next step should change.

A bounded retry can help with a temporary software failure. Repeating the same failed call indefinitely adds cost without adding evidence. A useful agent can stop, preserve completed work, and explain the specific obstacle that remains.

What have researchers demonstrated?

Published systems show how language models can coordinate scientific tools. Two examples help make the idea tangible:

ChemCrow, described in a 2024 Nature Machine Intelligence paper , connected a language model to 18 tools for chemistry tasks. The study included synthesis demonstrations and work on a new light-absorbing molecule, called a chromophore. It also reported that performance depended on the tools and how the agent used them. These are concrete demonstrations of tool coordination, with a defined experimental scope.

Coscientist, reported in Nature in 2023 , combined search, documentation, code execution, and laboratory automation for chemistry tasks. In one example, it consulted documentation to correct an invalid instrument command. The setup still required people to move plates manually. That detail helps explain what “autonomous” meant in the experiment.

Neither paper establishes that an agent can independently deliver an approved medicine. They demonstrate parts of scientific work that can be connected and tested. For drug discovery, the next questions concern performance on the actual task, repeatability, and the quality of the resulting evidence.

Autonomy is a set of permissions

It helps to describe autonomy with concrete actions. Can the agent read approved data? Run a calculation? Create a draft? Change a shared record? Request an experiment?

A team can allow the first three while requiring review for the others. Those boundaries should be enforced by the surrounding software. A sentence in a prompt is not the same as a tool permission.

The same principle applies to cost and time. A task can have a limit on model calls, compute use, or retries. When it reaches that limit, the agent should return its progress and unresolved questions. Stopping with a useful partial result can be better research support than continuing without a clear reason.

How to tell whether an agent helps

A fluent report is a poor measure of success on its own. Evaluate an agent on tasks a research team actually needs to complete, and include cases where the right response is to ask for more information or stop.

Useful measures include:

• Accuracy: Are identifiers, units, calculations, and source claims correct?
• Completion: Did it finish the agreed task and clearly identify unfinished work?
• Traceability: Can a scientist find the evidence behind a recommendation and reproduce the key calculation?
• Recovery: Does it handle missing data and failed tools without inventing a result?
• Practical value: After review and corrections, did it save time or improve the next decision?

Compare the result with a suitable baseline, such as an existing script, a fixed workflow, or the team’s usual process. Include the effort spent checking and repairing the output. More tool calls or a longer report do not necessarily mean more useful work.

AI agents at AgentCures

At AgentCures, we are building tools that connect molecular design, property assessment, experiment planning, and research documentation. Agents can help carry context between these tasks so that scientists spend less time searching, copying, and reformatting information.

Our approach keeps the work structured and version-controlled. Scientists need to see what changed, inspect the evidence, challenge assumptions, and decide which proposals deserve further work. A useful research record includes the questions that remain open.

For the wider scientific process, read our guide to how AI helps discover new drugs . Our articles on connecting AI with experiments and keeping research in version control explore the surrounding workflow.

The prospect is straightforward: less time stitching tools together, more time asking and testing good scientific questions. The value of an agent becomes clear when its work survives inspection and helps a team decide what to do next.