How coding agents run unattended here
A write-up of the harness that builds the work: a night runner, a review pipeline, a proof gate and a model router. Numbers are a snapshot taken 18 Oct 2026. This page does not update itself.
Budgets per ticket
Every ticket carries an est token budget and a model. Haiku takes mechanical edits and Sonnet takes work that needs judgment. The runner is budget-gated and stops on a limit error instead of retrying.
"Done" means the accept command passed and the work is committed
Each ticket ends with a command the runner executes. A run counts as done when that command passes and the work is committed. A failed command and an uncommitted finish both count as not done.
Accept-cmd: npx jest app/lib/evals/__tests__/gate.test.ts
Review before it lands
Some tickets go through a PR pipeline with an Opus review pass. It runs one review, one redo, then merge and a live check.
The proof gate
A Stop hook refuses to let an agent end a turn claiming it is done unless the turn quotes real output on a Verified: line or says what it did not verify. It has been running since 23 Sep 2026.
When work goes to a local model
Local Qwen, served by Ollama on a second machine, takes low-stakes drafting and checks. Claude takes anything that ships. The counter tracks delegated, completed, escalated, escalation rate, counted per task ID. A count of Claude calls avoided with no escalation rate would reward sending more work to the weaker model.
That is one tracked request, so this section shows the rule and the counter design. It does not show savings.
One place to watch it
A command center runs several agent chats in parallel. It shows live context size and token use for each project. The dashboard is private, so there is no screenshot.
The same rule applies to the product: scores gate the build.