zicato-design-proposer

Configuring a mutation proposer agent requires preserving contract inputs, enforcing model separation from the harness, and rolling epochs when behavior changes. This approach maintains evaluation integrity while generating grounded mutation experiments.

You need a custom ADK proposer when the built-in tool-using agent lacks the necessary skill-based reasoning instructions or curated read-only tools for your specific mutation-analysis needs. Customization allows distinct model separation and specialized workspace access.

Required evaluation guardrails for multi-agent mutation experiments include enforcing distinct proposer and harness models, hashing contract inputs, enforcing read-only workspace access, and rolling epochs when proposer behavior changes to prevent model collusion.

Roll epochs when proposer behavior changes by updating the proposer configuration and triggering a new evaluation cycle. This ensures that prompt evolution and mutation experiments are evaluated against a fresh baseline to maintain accurate harness metrics.

Configuring skill-based reasoning instructions does not compromise the evaluation harness if you enforce model separation, restrict the proposer to read-only workspace access, and preserve the proposer contract inputs. These guardrails maintain evaluation integrity.