Build a Self‑Improvement Loop for Your Skill

Recent discussions about using “loops” to drive agents often leave the question “what exactly is a loop?” unanswered. This article demonstrates a concrete, practical approach called a self‑improvement loop , where an agent continuously upgrades its own skill based on external feedback.

The core idea is illustrated with an issue‑classification skill that assigns new GitHub issues to categories such as “ready to implement”, “duplicate”, or “needs more info”. The skill can also be adapted for code‑review, vulnerability‑fixing, or incident‑response tasks.

Step 1: Build the Internal Agent Loop

The internal loop runs the classification skill whenever a new issue is created. It is implemented as a GitHub Action that triggers automatically, records the interaction, and can store logs in files, Slack, or GitHub comments.

The Action calls the cloud‑agent platform Warp and its Oz service, which fetches issue content, runs the classification skill, and adds appropriate labels.

Step 2: Build the External Self‑Improvement Loop

An external agent runs once a day, scans all classified issues, and detects any manual label changes made by reviewers. For example, if a reviewer changes a label from “ready to implement” to “needs more info” and explains the reason, the external loop records this feedback.

The external loop’s skill, executed by an encoding agent, generates a diff patch that incorporates the reviewer’s feedback and updates the classification skill. After merging the patch, the next run of the internal loop uses the improved skill, closing the feedback loop.

Step 3: Extend and Apply the Pattern

The same pattern can be applied to other skills such as code review, vulnerability remediation, or incident response. By storing the skill as a file and letting the external loop generate diff‑based updates, the system continuously self‑optimizes without manual re‑deployment.

All example code and workflow definitions are available in the linked GitHub repository, which demonstrates the full internal GitHub Action configuration and the external loop implementation.