Factory

The Agent Relay factory is an autonomous multi agent factory which integrates with any surface you want.

It runs the loop a human normally babysits. An issue is marked ready, and Factory picks it up, decides how to tackle it, puts implementers and a reviewer on it, opens a pull request, shepherds that PR until it is green, and moves the issue on. You stay in the loop where it matters and out of it where it doesn't.

The loop

1. Discover — poll or subscribe for issues that are ready and in scope.
2. Triage — decide the repo, the shape of the work, and how many agents it needs.
3. Dispatch — put implementers and a reviewer on it, in real checkouts.
4. Review — agents review each other's work before a human sees it.
5. Merge gate — the PR must be green and reviewed. Nothing merges by itself.
6. Transition — move the issue to human review, or to done.

Every stage is gated by the safety scope. Factory only touches work you have explicitly marked for it.

Triage decides the shape

Triage is the interesting step. Before any agent is spawned, Factory works out what the issue actually needs:

• Which repo — from labels, project, or keyword rules.
• How big — single for one implementer, workflow for a defined sequence, team for several agents across repos.
• Who — how many implementers (up to six), which reviewer, which model per role.

It tries a cheap heuristic pass first and only escalates to a model when the heuristic comes back thin or low-confidence, then merges the two. Most issues never need the expensive path.

If triage can't route an issue confidently, it says so rather than guessing.

See Issue labels and repository routing for the explicit agent:single, agent:team, and agent:workflow shapes and cross-repository dispatch.

Escalation

Agents ask instead of inventing. When an agent hits a decision it shouldn't make alone — an ambiguous requirement, a missing credential, a product call — it parks the work and posts to the issue thread, mentioning the stakeholders you configured. The thread is where the answer goes back in.

Nothing silently stalls: a stuck run surfaces rather than quietly failing.

Babysitting

Opening a PR is not the same as landing one. With babysitter mode on, a dedicated agent stays with each pull request and shepherds it: responding to review comments, resolving conflicts, and re-running until CI is green. When it's genuinely ready, the issue moves to human review.

You can also point it at an existing PR without turning the mode on:

factory babysit <PR>

Final review and merge always stay with a human.

Multi surface

Factory sits on top of Relayfile, so the surfaces it reads and writes are the full integration catalog — issues from Linear or GitHub, status to Slack, and anything else an adapter covers. Point it at the surfaces your team already uses.

Watching it work

Factory reports to the Agent Relay dashboard by default: every instance, every run, and a progress timeline of what each agent did — with a trace id correlating a run's whole life, and a closed privacy contract that has nowhere to put your prompts or source code. See Observability.

It also reports where you already are: lifecycle updates on the Linear issue, or GitHub comments plus factory:in-progress and factory:human-review labels, and a threaded Slack status per issue with your stakeholders mentioned when an agent needs a decision.

Start here