SHAPE: Learning to Search with an Evolving Harness for Target-Binding Antimicrobial Peptide Design

Designing deployable antimicrobial peptides (AMPs) is inherently multi-objective, requiring strong antimicrobial activity while balancing safety, AMP likelihood, and sequence diversity. Requiring effective engagement with biologically relevant targets, including drug-resistance and tissue-repair targets, adds another coupled objective and makes the search problem substantially harder. Recent computational methods provide strong peptide design priors, while agentic methods enable iterative search through LLM reasoning and tool feedback. Yet generative methods remain bounded by learned priors, and agentic search typically uses a fixed harness, so accumulated experience improves individual trajectories without improving the search system itself. This raises a broader challenge: how can a peptide discovery system improve its own search process as peptide search proceeds? We address this challenge with SHAPE, a multi-agent system coupling target-aware peptide search with harness evolution. To evaluate this setting, we introduce a target-aware AMP design benchmark that jointly measures predicted potency, safety, AMP likelihood, sequence diversity, and target binding. SHAPE distills search trajectories into executable condition-action operators, validates them by matched equal-budget comparisons against the current harness, and reuses them when applicable under a shared scorer budget. Across the benchmark, SHAPE improves the composite score by 23.99% over the strongest baseline under the same budget. Operator-guided edits achieve Pareto improvements in 42% of evaluated cases, versus 27% for LLM-reflection-driven edits. Evolved operators generalize across eleven generators: when frozen and reused for three rounds, they achieve a mean cumulative gain of 12.03% under an external composite score, with all sources finishing above their original means. Together, these results show that SHAPE improves both the designed peptides and the search harness itself by turning search experience into reusable search capabilities. To enable full reproducibility, we publicly release the SHAPE framework at https://anonymous.4open.science/r/SHAPE-AMP-Design-D44D.