📝Self-Improving Loops are agent systems that review their own results and revise their own prompts, skills, tools, or memory so that later runs perform better without manual rewriting.
A self-improving loop adds a reflection step to ordinary work. After a task, the agent or a separate reviewer agent examines what went wrong or ran slowly, then edits the artifacts that shaped the run: an instruction file, a 📝Skill, a tool, or a stored lesson. The next run starts from the improved version.
The approach builds on 📝Eval-Driven Loops, since improvement needs a signal to improve against. Without a clear evaluation, a loop can drift, overfit to recent cases, or accumulate contradictory rules. Practical systems therefore gate self-edits behind tests, human review, or version control so changes can be inspected and reverted.
Research treats self-improvement as an open frontier of agent design, covered in courses such as 📝Stanford CS329A: Self-Improving AI Agents. In production the common form is modest: agents that update their own guidance files and memories as they learn a codebase or a user's preferences.