method-evolution

To fix a recurring rubric violation, you evolve the agent's method by identifying the shared pattern, proposing a single concrete method change, and measuring the task-suite loss metric before and after the update. This ensures the issue stems from the method, not the tasks.

Agent method evolution is the process of updating an agent's core method artifact when the same rubric violation recurs across multiple tasks and domains. You should apply it when you observe a recurring pattern across at least five varied tasks, ensuring the problem originates from the method itself.

You measure the impact of a method change by calculating the delta loss (Δloss) across the task suite. Run the suite with the method_before and method_after configurations to compare the measurable loss metric, which guides auditable rollbacks if the change does not improve performance.

Yes, method evolution requires a task suite with at least five varied tasks. This minimum scale ensures the recurring rubric violations are detected across multiple domains, confirming the issue originates from the method rather than individual task anomalies.

The best way to document an agent method evolution commit is to name the observed violation pattern, propose a single concrete method change such as a rubric tweak or clearer prompt, and record the rationale alongside the measured Δloss to ensure a fully auditable evolution history.

Your agent keeps failing the same rubric criteria on different tasks because the underlying method artifact lacks the necessary constraints or clarity. When a violation recurs across multiple domains, it signals a method-level issue that requires a concrete method evolution rather than individual task fixes.