How it works

Who does what in a task: the scoper, tester, implementer, judge and reviewer, the two GitHub Apps beside them, and a task’s life.

When an agent’s pull request merges by itself, when it waits for your approval, and what you do when it hands off to you.

From a labelled issue to a merged PR: intake, triage, teams, revise, the merge gate and the kill switch.

The shared, append-only log of a task: the broker, the bridge, the web view, steering, room tools, approvals, and maku.

How one AgentRun becomes an isolated, attributed run: a gVisor sandbox on each cloud, per-run identity, octo-sts, branch and tag rulesets, and default-deny network policy.

Agent to agent: RunLore hands an incident to the factory, follows the fix, and checks after the merge that the alert is gone.

Moonshot (Kimi K3.1 Pro) and Mistral (Medium, Small) as selectable providers for public work: what they are, where data is processed, and how a class switches them on.

ai-gateway for humans and coding clients, the agent gateway for agent runs: identity, routing and what they share.

Where each request is routed, how to tune quality, cost and latency, what a run costs, and when self-hosting pays.

Turning the opt-in serving platform on, what it deploys, and its security posture.

The alternatives considered for each layer of the Agent Factory, why the current choices won, and what would make us reconsider.