Recursively Self-Improving Agents as Autonomous Software Engineering

How people use agents changes every day. Capabilities expand, use cases broaden, and the surface of workflows, stacks, codebases, and models keeps growing with them. To keep pace and ship the highest-performing product, the agent has to continuously learn without human intervention. At Factory, we close the loop between production behavior and improvements, and our agent Droid now ships its own fixes back into our codebase every day.

We will trace one full cycle of Droid improving itself, from detecting user friction in production to a Droid-authored PR, validated against our regression suites and merged. We will use the cycle to address key design questions: how to privately extract and cluster signal from sessions in aggregate, how to ensure quality as the agent and evals coevolve, and how to reduce human review burden as the loop scales.

The signal-to-fix loop is a general architecture pattern for autonomous software engineering: telemetry and logs for input, tests and evals for validation, merging for deployment, and monitoring for feedback. Any production AI system that can describe its own behavior, validate its own changes, and ship its own code compounds itself.