The AI-in-research story is often presented in two extreme forms. In one, the agent does everything, the scientist reviews the results, and hypothesis-to-data timelines fall from weeks to hours. In the other, machines cannot replace scientific judgment, so they offer no fundamental value beyond a spreadsheet. Neither view is accurate. Researchers making practical choices about what to use today are stuck in the space between these claims.
This note defines that current boundary as clearly as possible. Which research tasks do AI agents perform reliably today? Where do they fail, or produce results that need substantial expert interpretation? Capabilities change, so this is a current snapshot, not a permanent definition.
Where agents show clear value today
The clearest value of a well-designed research agent is coordination between established scientific methods. Methods such as docking, ADMET prediction, and MD simulation are understood. For a new question, the main friction usually lies in setup and data handling, rather than deciding which method to use or interpreting its results.
An agent can prepare the receptor by retrieving the PDB structure, removing solvent, adding hydrogens, and defining the grid. It can prepare ligands through 3D conformer generation, protonation assignment, and tautomer enumeration, submit jobs to a compute backend, and format the outputs. For an experienced computational chemist, this setup for a new target typically takes three to five hours. An agent doing it in ten to fifteen minutes removes coordination overhead, not scientific insight. That is a real contribution.
Agents also add value by applying pre-defined analytical pipelines consistently across large compound sets. If a team has agreed on pose checks, ADMET thresholds, and property cutoffs, applying those rules to 500 compounds is well suited to agent-based automation. Human reviewers can drift when applying the same criteria hundreds of times; agents apply them identically. That consistency matters even when the criteria are not optimal.
Where human judgment still does the real work
Current research agents do not genuinely understand the experiments they run. An agent that docks 200 compounds and returns a ranking has not judged whether the scoring function fits the target class, whether the protein is in a relevant conformation, or whether the leading compounds can be synthesized. It has carried out a procedure. Scientists must still decide whether that procedure fits the question and whether its results merit action.
Hypothesis generation is a field where overstating agent capability causes real harm. An agent can combine literature and suggest compounds supported by known structure-activity relationships, but this is pattern matching across published data, not mechanistic insight into why a particular protein may differ from its structural neighbors. If a team treats generated proposals as hypotheses rather than options for critical review, it may pursue suggestions that a knowledgeable medicinal chemist would immediately deprioritize. The issue is not that the agent generated them, but the role assigned to its suggestions.
Agents also underperform at anomaly interpretation. A docking run may show a known inactive scoring above known actives, or a similar compound series may split in predicted score. The agent can flag the anomaly, but it cannot determine whether the cause is a structural insight, a scoring artifact, or a data preparation error. That diagnosis requires expert review of poses, input structures, and interaction geometry. The review cannot be skipped.
The wet-lab handoff question
In wet-lab work, agents currently help with compound selection, identifying which members of a proposed series may be worth synthesizing before bench chemistry begins. They are not generally useful for experimental design. Questions such as "what assay format will give the most informative kinetics data for this interaction" or "should we use SPR or ITC for this binding characterization" remain best answered by scientists experienced in biophysical assay selection. An agent lacks an empirical model of what will work in a specific lab, with its instruments and a protein that may behave differently under particular buffer conditions.
The transition from computational prediction to experimental validation is not automated in practice, and likely should not be. A scientist reviewing docking results considers context unavailable to the agent, including the chemistry team's current synthetic capacity, assay availability, budget constraints, and project priorities. Automating the handoff would require automating that contextual integration. This is a coordination decision, not a computation problem.
How to set expectations before deploying an agent on a real question
Before using an agent to guide a real synthesis decision, spend time on a calibration set. Select five to ten compounds with existing experimental data, such as IC50, Kd, or binding binary results, and run them through the agent. Compare its docking scores and ADMET flags with the experimental evidence. If experimentally confirmed inactives all rank at the top of the docking list, the setup is flawed and the outputs cannot be trusted until the cause is diagnosed.
Calibration is not a criticism of the agent. It is the same validation required for any new analytical method before its results guide decisions. Skipping it because the interface is polished or the technology sounds credible is a process failure, not evidence of capability.
Agents configured for a specific target class and validated against known data can provide a reliable, time-saving addition to research workflows. Scientists remain responsible for staying within the validated scope and escalating unexpected results for expert review. The agent does not know when its outputs should be distrusted. That judgment remains with the research team.
