Agents v0.4.0 is focused on making the intelligence catalog easier to evolve safely.
This release introduces versioning for prompts, skills, and guardrails, together with a self-improvement workflow driven by maintainer feedback.
The core idea is simple: when an agent makes a mistake, the fix should not only apply to the current PR. The feedback should also become a signal for improving the prompts, skills, or guardrails that shaped the agent behavior.
Maintainers can now leave /agents improve comments on GitHub issues, PRs, or review comments. Agents ingests that feedback, resolves the surrounding attribution context, runs an internal intelligence analyst, and presents an editable proposal for changing the relevant catalog asset.
If the analyst has enough context, it prepares a proposal bundle that can update prompts, skills, or guardrails. If the feedback is too ambiguous, it asks for clarification instead of guessing.
This creates a tighter loop between real review feedback and the intelligence catalog that powers the fleet. It helps reduce repeated review noise, makes prompt evolution more concrete, and gives teams a structured place to improve agent behavior over time.
Documentation for the flow is available here:
https://github.com/eloylp/agents/blob/main/docs/self-improvement-feedback.md