Notes from building a software factory.
A personal tech blog about agent-driven development, harness design, and the small services that keep a coding-agent factory honest.
9 entries
The gates, not the autonomy
Living with patchrelay, the autonomy turned out to be the part I reach for least deliberately. What actually changed how I ship are two gates — a reviewer and a merge queue — that do not care whether a human or an agent wrote the code.
Review volume is the wrong bottleneck
PatchRelay can keep agents busy, and that quickly makes manual review feel impossible. The review-quill repair loop absorbs much of the code-review work. The harder question — whether the agents are still steering the product in the right direction — is still open.
Don't pay for the same review twice
Every review-quill round costs time and tokens. Two patterns waste them — parallel agents whose PRs should have been sequenced, and clean rebases that produce a new SHA without changing the diff. The factory plans to avoid the first and uses `patch-id` to skip the second.
Merge trees: a clean mental model
A PatchRelay model for changes, reviews, and landings, built from primitives Git already gives you: commit trees, patch-id, merge-tree, and fast-forward landing.
From YOLO to patchrelay
Notes on a year of agent-driven development — permission prompts, a rented Hetzner box with nothing on it, a parallel-agent experiment that turned into merge-conflict hell, and the small annoyance that became patchrelay.
patchrelay: Linear issues in, pull requests out
patchrelay is the deterministic orchestration layer that turns a delegated Linear issue into a linked pull request and keeps that PR healthy until merge or close.
merge-steward: speculative integration, parallel validation, fast-forward landing
merge-steward is a self-hosted merge queue with three decisions — test every PR on `main + diff`, validate cumulative speculative chains in parallel, and fast-forward `main` to the exact tree CI ran against.
review-quill: a strict reviewer for your coding agent
Coding agents focus on the task and forget the surroundings: docs drift, tests go stale, sibling files keep old assumptions. review-quill is the strict reviewer that keeps sending the PR back until the repo is aligned again.
Picking an agent harness when the SDK terms are murky
The five real options for embedding a coding agent into a custom factory, why the SDK licensing question pushed me to the Codex App Server, and why session-attach belongs in the agent runtime, not the terminal layer.