Why Science Needs Agentic AI, Not Just Predictive AI

After two weeks of enzyme kinetics assays in early 2025, a small research group had predictions for over 400 candidate conditions from a model with reasonable accuracy. Their lab was still running four experiments per week, exactly as before. The model could answer supplied questions, but it did not help determine which questions deserved to be asked next.

The distance between forecasting an outcome and choosing the next test marks the boundary between predictive AI and agentic AI. That distinction changes what AI can contribute to the scientific workflow.

Two Questions, Not One

A predictive model asks: given these conditions, what result is likely? That is useful. Estimating protein binding affinity for a proposed compound can prevent wasted synthesis. Predicting reaction yield can narrow the search before chemistry begins. Prediction is substantive work, and it removes poor options early.

Research, however, turns on a different question: which conditions should we test next? That is a decision problem. Prediction maps input conditions to expected outcomes. Experiment selection maps the current knowledge state, including known results, uncertainty, and experiment cost, to a ranked set of actions.

Treating prediction and selection as one capability has stalled many AI-for-science tools. Accurate predictions look like progress, so the failure can be hard to see. A lab gets a model that answers queries well, yet experiment cycles do not accelerate and the information frontier barely moves. Prediction becomes an input to the old decision process, rather than removing its slowest steps.

The researchers who shared that feedback in 2025 were using their model as intended. The issue was conceptual: generating a prediction for every candidate and deciding what to run are separate activities. Only the latter advances the experimental cycle.

What Agentic Means

"Agentic" is often used loosely in product marketing, so we want a precise definition. In research automation, an agentic system acts in the experimental world. It reads the current knowledge state, chooses the next experiment, submits it, receives the result, and updates its model of the system. It repeats this cycle without asking a researcher to manually translate each handoff.

This is not a chatbot discussing strategy, nor a model that produces a ranked list for someone to re-enter into a LIMS or instrument controller. Agency closes the loop.

Three capabilities make this possible. First is a surrogate model of the scientific system, which predicts candidate outcomes. Second is an acquisition policy that scores experiments by expected information gain, expected improvement over the current best, or a related criterion. Third is an execution layer that submits work to instruments or simulators and returns structured results. Prediction is one part of the architecture, not the whole architecture.

An Information-Theoretic View

Suppose a surrogate reaches 80% accuracy across a 10,000-point parameter space. You can estimate the outcome at each of those 10,000 conditions, but you can run only 50 experiments. Which 50 should you choose?

A random sample spreads 50 measurements without considering uncertainty or where the optimum may lie. A purely exploitative policy repeatedly selects conditions predicted to be best, clustering 50 runs near the current leader while often missing global optima and learning little about the wider landscape. Expected improvement and upper confidence bounds instead balance exploitation of strong regions with exploration of uncertain ones.

The acquisition function is the intellectual center of an agentic science system. It is a decision rule, not a prediction. Most research AI tools omit it and leave experiment selection to the scientist while offering prediction as a service. Without acquisition, the researcher still performs the loop's hardest cognitive step manually.

The Scientist's Role Remains

It is reasonable to worry that agentic AI removes scientists from the loop. We think a well-designed system does the opposite. A researcher consumed by "what should I run next, given this result" has less attention for defining meaningful results, connecting observations to theory, or asking what an unexpected finding says about mechanism.

Agentic systems should handle mechanical iteration: propose, run, observe, update, then propose again. Scientists should retain the judgments that matter: setting objectives, interpreting anomalies, recognizing when the search space must change, and generating hypotheses an automated loop cannot produce alone. The loop can navigate a defined space efficiently, but it cannot reliably recognize that the space itself needs redefining.

That boundary is important. Discovery Loop does not replace scientific judgment; it removes operational iteration that does not require it. Confusing those tasks makes "AI replacing scientists" sound broader than the actual, narrower proposition.

Where Prediction Still Belongs

This is not a claim that predictive AI lacks value in research. Many discovery steps call for prediction rather than decision-making. Pre-screening a large library is one example: filter 100,000 candidates to 500 worth placing in the acquisition queue. Virtual screening in drug discovery, machine-learned interatomic potentials for materials simulation, and generative chemistry models perform forward prediction effectively, and their outputs can feed acquisition policies.

The problem starts when prediction is presented as the finished product. "Our model predicts yield with 85% accuracy" does not answer a lab deciding which 20 experiments to run next week. Prediction must sit inside a decision architecture that changes action, not merely knowledge.

Discovery Loop was not built around a better predictor. We built a system that accepts an existing prediction capability, including an external model a lab already trusts, and places it inside an acquisition loop that selects, queues, reads, and updates automatically. Prediction supplies a component; the loop changes cycle time.

The labs seeing the greatest benefit have shifted from asking "what will this condition produce" to asking "what does the loop recommend I run tomorrow." The change in question framing is the practical point.