Signals: Toward a Self-Improving Agent

Factory's Signals turns recurring problems in coding-agent sessions into proposed software fixes. According to the company, language models review sessions daily for signs of struggle such as error events, repeated rephrasing, rejected tool calls and backtracking. Descriptions of these friction moments are clustered so the system can identify recurring problems. Signals account.

When a pattern's frequency crosses a threshold, Signals files an issue that Droid takes through implementation and code review. A person still approves the proposed change before it is merged; the article says the process is not fully automated yet. The loop therefore connects product usage to repairs in the agent's software, with human approval at the final step. Factory gives tool timeouts and truncated output as examples of problems that led to fixes. Closing the loop.