A self-healing codebase built on compounding agent loops

Problem

Keeping a codebase healthy — finding and fixing the steady stream of bugs — is constant toil, and there's a lot of hype around 'compounding loops' but little on what actually works in practice.

Approach

Millin Gabani built the loop on top of Worklayer, a 'private environment for agents' (apps, data, skills, and memory exposed via MCP and CLI); because the product is itself agent-used, the dev MCP server can be connected to coding agents. The self-healing loop runs in two phases — agents find issues, then fix and ship them — across four apps: Notion as an agent-managed issue system-of-record (with views the agent created); PostHog for high-cardinality event instrumentation (following Boris Tane's 'logging sucks' skill); Traces (a CLI) to collect rich QA/MCP trace data the agent can search; and Graphite for stacked, incremental PRs reviewed by Devin. Two daily Codex automations (find-issues and fix-issues) are each backed by a skill. Humans are limited to triage — assigning each issue to agent, user, or discard — via a custom keyboard-driven interface Millin built on top of Notion in about ten minutes by dictating requirements. Fixes flow GitHub → Google Cloud Run.

Results

Two founders built the entire product in three months on these compounding loops, and the system finds and fixes issues daily. It runs inside a single $200 Max plan when the skills are optimized, and the loops are easy to repair conversationally — Millin fixed a broken run (an MCP name casing bug) by teaching the agent and updating the skill, since automations pull instructions from skills rather than the prompt. The custom triage interface has been in daily use for about a month and a half.