Agentic AI

What makes AI “agentic”?

An agentic system typically combines four capabilities:

• Planning: breaking a broad goal into smaller steps.
• Tool use: calling external resources, such as a database, search engine, or lab equipment.
• Reasoning from feedback: interpreting the results of its actions, including messy, noisy real-world feedback, and adjusting its next step.
• Memory and context: retaining relevant information across steps.

This is a shift from earlier AI systems, which generally did one thing well and needed a human to interpret the output and decide what to do next.

How is this different from a Large Language Model on its own?

A Large Language Model (LLM) is often the "reasoning engine" at the core of an agentic system, but an LLM by itself is typically a single-turn tool: it receives a prompt and generates a response. An agent built around an LLM adds a layer of orchestration on top: the ability to break a task into steps, decide which tools to call at each step, check its own outputs, and keep going until the broader goal is met, rather than stopping after one exchange.

How does an agent define its task?

An LLM doesn’t decide on its own what job it is doing. The task starts with a goal stated by a person, usually as a prompt. The orchestration layer then wraps that goal in further instructions: the role the model should play, the rules it should follow, and the tools it is allowed to use. From there, the LLM proposes a plan and the first step, the orchestration layer runs it, and the result is fed back in so the model can decide on the next one.

In other words, the task is defined jointly. The person supplies the goal, and the orchestration layer and the LLM turn it into a sequence of concrete steps. Where a goal leaves details unspecified, such as which data to use or which method to apply, what fills that gap depends heavily on how the orchestration layer is built.

Two agentic systems can be built on the exact same underlying LLM and still behave very differently, because so much of an agent's actual behaviour depends on this orchestration layer: how it defines its task, which tools it has access to, and how consistently it makes the small decisions a task doesn't explicitly specify.

How does an agent know when to stop?

An LLM has no built-in sense of “finished.” In practice, an agent stops for one of a few reasons:

• The goal appears to be met: the model produces a final answer instead of requesting another step.
• A limit is reached: the orchestration layer caps the number of steps, the time, or the cost.
• A check is passed: the output is tested against defined criteria before the process ends.
• A human is needed: the agent pauses to ask for approval or clarification.

Defining this ‘stop point’ is key for correct agentic functioning. An agent that stops too early returns an incomplete answer, while one that keeps going can loop, or build further steps on top of an earlier mistake.

What’s an example of an orchestration layer for an agent?

K Pro, Owkin’s agentic AI platform for biopharma R&D, is a useful example. It orchestrates a frontier LLM, so the model provides the reasoning, while the layer around it (also referred to as the “harness”) decides what happens in what order, which purpose-built tools do the work, and how domain conventions are applied.

Consider a routine request: “in stomach cancer data from TCGA, classify patients as MUC16-mutant or wild-type, then compare tumour mutational burden between the two groups.” The question sounds complete, but it leaves several decisions open, including which patients have usable mutation data, what counts as “mutant,” and which statistical test to run.

K Pro’s orchestration ensured that this request was routed through tools built and tested for each step. It also ensured those decisions were applied the same way every time. Compare that to the same underlying model without the K Pro harness, which produced different cohort sizes across repeat runs, while reaching the same overall conclusion. This shows how the orchestration layer makes decisions that affect a model’s output. You can read more about that full comparison in How Does K Pro Compare to Claude?

 How is agentic AI being used in healthcare and biomedical research?

Agentic AI is increasingly being explored for tasks that resemble an open-ended research process, rather than a single prediction, for example:

• Scientific discovery: an agent proposing hypotheses, selecting experiments, interpreting lab results, and adapting its strategy over multiple rounds, sometimes referred to as an "AI scientist."
• Data analysis workflows: an agent that can independently query a dataset, apply the appropriate statistical method, and produce a structured analysis, rather than requiring a human to manually run each step.
• Clinical and research decision support: an agent capable of pulling together information from multiple sources (patient records, trial data, published literature) to support a specific research or clinical question.

What are the risks or challenges specific to agentic AI?

Agentic systems introduce challenges that go beyond those of a single-task model:

• Reliability and reproducibility: because an agent makes many small decisions along the way, two runs of the same task can take different paths and reach different conclusions, even when using the same underlying model. This makes benchmarking agentic systems significantly harder than benchmarking a narrow, single-output model.
• Error compounding: a mistake made early in a multi-step process can carry through and affect every subsequent step, rather than being an isolated, one-off error. For example, hallucinations made by the system early on in an agent’s steps can build.
• Reduced interpretability: with more steps, more tool calls, and more autonomous decisions involved, it can be harder to trace exactly why an agent arrived at a particular conclusion, compared to a single-step model's output. This connects directly to the broader challenge of interpretability.

Does agentic AI replace human researchers or clinicians?

Not currently, and not in the way the term "agent" might suggest. Agentic AI systems are generally best understood as a way of automating and accelerating specific parts of a research or analytical workflow, rather than as fully autonomous replacements for human judgement, particularly in high-stakes clinical contexts.

What can agents not do?

Agents are limited in several important ways. They cannot currently:

• Decide which questions are worth asking: an agent can pursue a goal, but choosing which scientific or clinical problem matters is a human judgement.
• Take accountability: an agent cannot defend its output to a regulator, a review committee, or a patient. A person still has to stand behind the result.
• Replace real-world validation: a hypothesis generated by an agent is not a result. It still needs experimental or clinical evidence, which can take years to arrive.
• Reliably flag their own unstated choices: an agent can return a confident, well-formatted answer to an analysis nobody deliberately chose, with nothing on the page revealing it.
• Replicate human care: as with other AI, an agent can mimic empathy but cannot truly understand it, so it is not a substitute for the human relationships at the heart of clinical care.

Can you give an example of humans and agents working together?

A computational biologist working on target discovery might ask an agent such as K Pro a question about how a gene’s expression relates to patient survival. The scientist sets the question and the conventions that matter, such as how the cohort should be defined. The agent then carries out the multi-step analysis: selecting the patients, running purpose-built tools, and producing the results. The scientist reviews the route the agent took, checks that the cohort and methods make sense, and decides whether the finding is strong enough to take to a team or committee.

The agent saves the time spent on running each step by hand, while the human keeps the decisions that need judgement and accountability.

Given the reliability and interpretability challenges above, human oversight remains essential, especially for anything that will ultimately inform a clinical decision or a significant research conclusion. In practice, this often means a human reviewing and verifying an agent's output, or its reasoning path, before acting on it.