DrugHarness: How an Evolvable AI Agent Accelerates Drug Discovery

Computational drug discovery is shifting from "getting a model to run" toward "organizing a reusable set of scientific operations." The recently proposed DrugHarness couples composable scientific execution with conditional skill induction, attempting to make AI agents get stronger the more they are used in real drug design tasks.

From "Tool Calling" to "Skill Evolution"

The core idea of DrugHarness is not to pile more tools onto the agent, but to let it distill reusable "skills" from execution trajectories. Its environment integrates more than 80 scientific operations, biomedical resources, and task-specific code, while preserving the associations between candidate molecules, results, and artifacts. After the agent completes a task, the system induces action hypotheses with observable applicability conditions from the trajectories and result feedback. These hypotheses then undergo source verification or targeted execution trials by an independent review agent, and only those that pass are admitted as skills.

This "conditional skill induction" is fundamentally different from ordinary prompt reuse: skills carry explicit applicability boundaries and supporting evidence, and are revised as subsequent feedback arrives. In other words, the agent does not memorize "how it was done last time" but learns "under what conditions what should be done."

Conditional skill induction is the core mechanism of the DrugHarness method, analogous to "option discovery" in reinforcement learning and "abstraction extraction" in program synthesis. Traditional agent tool calling is one-off: each task is planned from scratch, and historical experience is only passively reused as conversation context or retrieved snippets. Conditional skill induction instead binds action sequences from successful trajectories to trigger conditions (such as molecular scaffold features, target types, and constraint types), forming reusable units of the form "if condition C is met, then execute action sequence A," accompanied by supporting evidence. The key differences are threefold: skills have explicit applicability boundaries, and use beyond those boundaries is intercepted by the review mechanism; skills have evidence provenance, sourced from source verification or targeted trials; and skills can be revised or even discarded based on subsequent feedback. This makes the skill library a continuously evolving knowledge asset rather than a static set of prompts.

MolBench and MolClaw-CC require background context to understand the benchmark comparison. MolBench is a benchmark for molecular property prediction and affinity comparison, typically including target-ligand pairing and activity ranking tasks; MolClaw-CC is a previously reported molecular agent baseline characterized by tool calling and context compression. In this article, "affinity comparison accuracy" refers to the agent's correctness rate in judging which of two candidate molecules has higher binding affinity for the same target, and "top-three active hit count" refers to the number of molecules among the top three in the recommendation list that truly have high activity. DrugHarness outperforms by 10.8 percentage points and 55% even without evolution enabled, indicating that its environment integration and task organization already bring gains on their own, with skill evolution further amplifying them on that basis.

Benchmark Tests: Already Beating Baselines Without Evolution, Significant Gains With It

On MolBench and SMDD-Bench, DrugHarness demonstrated capabilities in molecular screening, optimization, and constrained design. According to the paper's data, without evolution enabled, its affinity comparison accuracy on MolBench already exceeded the previously reported MolClaw-CC results by 10.8 percentage points, with average top-three active hit count higher by 55%.

With skill evolution enabled, the improvement is more pronounced: retrospective evaluation on skill development tasks shows Qwen's affinity comparison accuracy rising from 62.2% to 81.1%, and optimization success rate from 79.5% to 87.2%. This shows that skill evolution is not merely icing on the cake but provides substantive amplification of model capability.

Prospective Validation: From Computational Design to Experimental Loop

More convincing is the prospective HPK1 study. Among system-specific candidate molecules, DrugHarness achieved more favorable median docking scores than Biomni. More critically, one synthesized compound achieved a cellular IC50 of 14.19 nM, a 2.64-fold potency improvement over the best input reference molecule, with activity comparable to selected compounds designed by two senior medicinal chemists.

The significance of this result lies in closing the loop from "computational design to experimental validation." Most computational drug discovery work stops at scoring and ranking, whereas DrugHarness's outputs go directly into synthesis and cellular experiments, yielding quantifiable potency data.

HPK1 and cellular IC50 are key to understanding the weight of the prospective validation. HPK1 (hematopoietic progenitor kinase 1) is a negative regulator of the T-cell receptor signaling pathway; inhibiting HPK1 can enhance anti-tumor immune responses, making it a hot target in immuno-oncology. Cellular IC50 refers to the compound concentration required to inhibit kinase activity or related signaling by 50% at the cellular level; lower values indicate stronger potency, and 14.19 nM falls in the sub-micromolar high-activity range. The "median docking score" in the article comes from molecular docking, a computational method that uses scoring functions to estimate the stability of ligand-target binding conformations; more negative scores generally indicate more favorable binding, but it is only a computational prediction and cannot replace experiments. Chaining docking ranking, synthesis, and cellular experiments into a closed loop is precisely the core evidence distinguishing this work from purely computational screening.

The independent review agent design deserves separate mention. Letting an agent induce skills from its own trajectories carries the risk of "self-confirmation bias": erroneous hypotheses may become entrenched through repeated citation. DrugHarness introduces a review agent separate from the skill inducer, setting admission thresholds through source verification (checking whether the data or code sources a skill relies on are genuine and reliable) or targeted execution trials (actually running the skill under controlled conditions and observing results). This is equivalent to adding a quality gate to the skill library, making skill evolution not unconstrained self-reinforcement but a screening process with evidentiary thresholds.

Implications for Scientific Agents

DrugHarness's experiments support a judgment: the competitiveness of scientific agents depends not only on the underlying model but more on whether auditable, revisable skills can be accumulated within the task flow. The independent review agent design also addresses a common concern—whether letting agents self-evolve will accumulate errors. By setting admission thresholds through source verification and targeted execution trials, the quality of the skill library is controlled.

Of course, the cost of skill evolution is the additional overhead of review and trials, and its scalability in larger task spaces remains to be observed. But based on current evidence, conditional skill evolution is a viable path connecting computational drug design and experimental validation.

One-line Summary

DrugHarness uses conditional skill evolution to make drug discovery agents stronger with use and closes the loop from computation to experiment.

Paragraph Summary

DrugHarness proposes coupling composable scientific execution with conditional skill induction, enabling AI agents to distill skills with applicability boundaries and evidence support from execution trajectories in drug design tasks, which are admitted to the library after validation by an independent review agent. The environment integrates more than 80 scientific operations and biomedical resources. On MolBench, without evolution it already exceeds MolClaw-CC by 10.8 percentage points, with top-three active hit count 55% higher; with evolution enabled, Qwen's affinity comparison accuracy rises from 62.2% to 81.1%, and optimization success rate from 79.5% to 87.2%. In the prospective HPK1 study, a synthesized compound achieved a cellular IC50 of 14.19 nM, a 2.64-fold potency improvement, comparable to compounds designed by senior medicinal chemists, closing the loop from computational design to experimental validation.