CLARA: Compound Library Active Learning with Reasoning Agents

Batch active learning for ligand discovery requires a recurring decision about whether to exploit high predicted affinity, reduce model uncertainty, or broaden chemical coverage. We test whether a language model can make this strategy-level decision while candidate ranking and label revelation remain controlled by a deterministic Gaussian-process environment. Across four protein targets, ten paired seeds, and a 360-compound assay budget, a GPT-5.6 Sol agent outperformed random sampling in all 40 matched campaigns. Its final top-2\% recall was 0.412 on D2R, 0.374 on Mpro, 0.448 on TYK2, and 0.683 on USP7. It was statistically tied with the strongest fixed policy on Mpro and USP7 but underperformed target-specific leaders on D2R and TYK2. The agent shifted from exploration-oriented early batches toward exploitation and produced 686 testable threshold claims with an 86.3\% atomic success rate, although 19.8\% of forecasts were not numerically scorable. In 200 prespecified information-ablation campaigns, removing uncertainty, diversity, or memory caused only small endpoint changes, while shuffling explanatory evidence reproduced the full allocation path in 39 of 40 paired campaigns. A deterministic 80\% availability constraint reduced recall on every target but preserved the target-dependent pattern of wins and losses. These results support language-model strategy control as a feasible and auditable layer.