The evolution of AI: From chatbots to agents

Everyone has experimented with use-case 101, chatbots, and attention moves to expediting more laborious human workflows. We have seen the rollout of numerous efforts from OpenAI that provide more than just ‘chat’ functionality. For example, o1 released in 2024, was capable of advanced reasoning, meaning it could think through complex problems and generate a chain of thought before providing an answer.

Fast forward to 2025, The Year of The Agents. From OpenAI, we have already seen the Operator, an agent for completing web-based tasks, and now Deep Research, an agentic capability that conducts multi-step research on the internet for complex tasks. The Year of Agents can be interpreted as the year people start extracting know-how from human subject matter experts (SMEs) and codifying this into their software. But where could this agentic AI bring value to the life sciences? And, importantly, what considerations should be made when looking at agentic AI?

What is agentic AI?

When it comes to talking about agentic AI, it is often good to start with what we mean by agency, defined as the below:

We can extend the above to define AI agents, utilizing AI in a process or task where it is given some level of agency, i.e., the freedom to make decisions on behalf of a user who has delegated such responsibility. Agents are designed in such a fashion that they can complete specific tasks that may well take multiple steps to complete manually, such as target prioritization. A target prioritization agent, for example, will take a therapeutic area as input (e.g., Type 2 Diabetes) and provide a list of prioritized potential therapeutic targets (e.g., GLP1R ).

It makes sense to give agents the right tools or functions to complete the job.

For example, the target prioritization agent, which takes a disease area as input, will need to access:

1. a database describing disease-gene relationships
2. a database capturing the druggability of targets
3. literature sources for providing evidence or novel hypothesis

Considerations when it comes to agentic AI

Before developing any solution, be sure to truly understand the problem you are trying to solve. Refrain from trying to identify problems for which to apply your solution.

For outcomes of agents to be trusted, traceability, and therefore transparency are key. It is not enough to say we are focussing on target X as part of our pipeline because agent Y told us to. It is vital to be able to understand WHY decisions were made – what tools were used, what data was reviewed, and what was the reasoning that led to the output from the agent?

For agents to consider the relevant search space when addressing a task, it is essential that they can access all necessary resources. This comes in the form of documented APIs presented to the agents as tools. Not only should resources be accessible with clear documentation, but there should also be some level of tool ‘description’ that can be interpreted, for example, “Need to know what the mechanism of action of a drug is? Then go to this service?”. One should also consider permissions of the data that is accessed by the agent.

Although accessibility to resources is critical, outputs are only as good as the quality of the data that they have access to. This comes from having gold standard data for the domains in which you play.

LLMs are great at understanding language but can struggle to understand the synonymity and ambiguity that is ever-present in science; science is more complex than language.

For example, knowing that a mention of TA123 is the internal code for the gene GLP1R, is crucial when it comes to harmonizing data and amplifying pertinent signals.

Retrieving data from a single source becomes problematic without understanding the varied ways in which an entity can be referred to; for example, most open-source databases don’t have great semantic search and instead rely on basic synonyms. An agent will search with perhaps one synonym, e.g., Type 2 Diabetes, it will then assess results and, if it’s not satisfied (because for example, the results don’t have the right context), it may try other synonyms such as T2D and iterate around again – a slow and costly process. If, instead, the search tool was able to better deal with synonyms and higher-level concepts upfront, the search can be made a whole lot more efficient.

Furthermore, different resources can represent their data in multiple ways, both syntactically (e.g., unstructured text in search tools vs structured data in graph databases) as well as semantically (a search system may use MeSH to capture diseases, whereas your graph index may use the DOID) – having the ability to describe equivalency via mappings (e.g., MeSH:D003924 and DOID:9352 are both identifiers for the disease Type 2 Diabetes Mellitus).

By providing an LLM that understands language with scientific knowledge captured in ontologies, human explainability of outcomes can be provided as part of the transparency step.

Be careful how much autonomy you give your agent. Certain tasks will require little and may be better suited to a more linear process defined upfront. As opposed to trusting an LLM to decide the next step to take in a process, give it a clear plan, or even reach out to an SME for input and review before moving on to the next step in the execution plan.

There are scenarios where autonomy may be better suited, such as identifying novelty, but again, be sure to use your SMEs for evaluation.

By focusing on having a clear set of tools or functions available, one can build a multitude of different agents atop. Think modular and document these as best you can. It is important to think bottom-up when developing agents.

Ensure that strict guard rails are in place when building out agents, and make it clear what the agent is not to do as much as it is to do!

A complex technical setup does not equal fit for purpose. It is vital that the subject matter expert remains a key part of the development of any solution, in terms of design and critically, in their evaluation.

How can ontologies support agentic AI?

Agents often have access to multiple tools. These tools may well contain data that is captured in multiple technical or syntactic formats, such as an SQL database, a knowledge graph, or a document index accessible via an API. While LLMs are fantastic at translating language, for example, natural language to query language Y, such as SQL, cypher, or a RESTful API call (given the relevant documentation), they can struggle to convert natural language to the relevant identifiers used to describe the things of relevance within these systems.

Furthermore, in scenarios where different systems use different standards to represent the same things, it is vital to be able to harmonize the output from one system with the output of another, that is, to understand equivalence.

Ontologies can help alleviate all the issues highlighted above. Enriched versions of them can be used to convert natural language to identifiers, converting unstructured text to machine-readable identifiers when building databases. The same approach can be used to convert the string Glucagon-like peptide-1 to the ID GLP-1 for querying over a knowledge graph. Furthermore, ontologies can be used to capture equivalence where multiple standards exist to define the same domain, such as diseases in MeSH, DOID, ORDO to name a few.

The synergy between quality data, ontologies, and GenAI in the context of agentic AI provides a powerful toolkit that can be used in combination to provide time-saving solutions to the drug discovery process.