Coming soon

Hand it a ticket. Get a pull request.

Autopilot for your backlog. Agent Factory runs coding agents through our built-in workflows, or ones you write yourself. It plans, builds, runs your checks, reviews the diff and opens the pull request. You just merge.

Agent Factory job page: a bug report taken through 12 runs — questions, research, design, outline, implement, gate, review, summarize, push and open PR — finished with PR #23 and a total cost of $3.72.
12runs
10stages
$3.72total cost
2,745tests passed at the gate
#23pull request, opened
How it works

One issue in. A reviewed pull request out.

Stop babysitting one agent in one terminal. Agent Factory runs the whole workflow, one recorded run per stage, and stops only where you want a say.

  1. 01

    Label an issue

    GitHub Issues first, then Jira and Linear. Or create a job in the Agent Factory app. A worker on your machine picks it up.

  2. 02

    Every stage, its own run

    Questions, research, design, outline, implement. Each is a separate coding-agent run you can watch live. Approvals wait for you.

  3. 03

    Checked, then reviewed

    A gate runs your repository's own checks, like pnpm gates:all. Failures loop back to a fix. A review stage reads the diff.

  4. 04

    Get a pull request. You merge.

    It writes the PR summary, pushes a branch and opens the pull request. If CI goes red, it fixes and pushes again.

Job page: status in the Path rows sits lower than the rest of the row
agent-factory · rpi · Research, plan, implement
    Running… PR #23 total $0.00
    Features

    Built for work you can check.

    Every screen below is the real product, captured from a live factory. Nothing redrawn.

    01 · Runs

    Watch every run, live.

    Each stage is its own coding-agent run with a streaming transcript, turns, cost against its budget, tokens and the diff it made.

    An implement run: the live transcript of reads, edits and bash commands, with a summary of cost $0.56 of $5.00 and 23 of 60 turns.
    02 · Gates

    Loops until green.

    A gate runs your repository's checks. A failure routes to a fix stage, and the gate runs again.

    Path: implement, gate fail (3 of 4 steps passed), fix, gate pass (4 of 4 steps passed), finished.
    03 · Recovery

    A lost run, retried.

    If a worker stops heartbeating, its lease expires and the run is queued again. A crashed machine doesn't stall the job.

    Banner: Lease expired — the worker stopped heartbeating 14m 20s into the run, so the control plane released the lease. Attempt 6 has already been queued.
    04 · Approvals

    Push back. It listens.

    Approval stages wait for you. Request changes with a comment, and the stage runs again with it.

    Path rows where approve-design requested changes with a comment, and design ran a second and third visit before approval.
    05 · Guardrails

    Guardrails built in.

    Read-only stages are enforced. Each command gets its tools and Bash patterns allow-listed. Budgets cap cost, time and loops.

    The implement command: max turns 40, budget $3.00, timeout 20 min, allowed tools Read, Glob, Grep, Edit, Write, and Bash only for pnpm test, typecheck, ledger, git diff and git status.
    06 · Artifacts

    Every draft, versioned.

    Questions, research, designs and outlines land as artifacts, and each revision is kept with the attempt that wrote it.

    The Artifacts tab: 03-design-design.md with a version menu listing v3, v2 and v1 and the attempts that wrote them.
    07 · Your process

    Your process, in the repo.

    Start from a built-in workflow or write your own in .agent-factory/, next to your code. Every job runs the same stages, checks and review, whoever starts it. No vibe coding.

    A repository's workflows table: presets and files under .agent-factory/workflows/.
    08 · Ledger

    Every run, on the record.

    Status, reason, duration, turns, cost and changes for every attempt. The lost one included.

    The Attempts tab: 12 runs with status, reason (run 5 lost, lease_expired), duration, turns and cost, totalling 34m 02s and $3.72.

    Runs on your machine

    Workers lease jobs and run agents where your code already lives.

    Budgets that stop

    Caps on cost, turns, time and loops. A job that hits one ends.

    Human in the loop

    Approve, request changes or stop at any stage you choose.

    Keyboard first

    Jump to any job or run with ⌘K.

    Proof

    It shipped a fix to its own UI.

    We handed the factory a bug in its own job page: status badges sat a few pixels low. Twelve runs later, the pull request was open, checked and reviewed. Every autopilot job runs like this one.

    • Runs, across 10 stages 12
    • Gate, pnpm gates:all 2,745 tests passed
    • Review verdict approve · 0 findings
    • Lost run, recovered #5 → retried as #6
    • Total cost $3.72
    GitHub pull request #23, 'Center Path row items to fix status baseline misalignment', merged, opened by agent-factory for job 46fd303c.
    Roadmap · Not built yet

    What's next.

    Intake and CI babysitting land before the first invites. Waitlist members hear first.

    Next · Intake

    Label an issue. Get a PR.

    GitHub Issues first, then Jira and Linear. No need to open the app. The factory reads the issue, opens a draft PR and comments back with links.

    Next · Babysit

    Red CI? Fix pushed.

    It watches the PR. When CI fails, it fixes, gates and pushes again. You still merge.

    Next · Isolation

    Each job, its own container.

    Own ports, own database, own model login. Your keys stay on the host.

    Next · Upkeep

    It finds the work too.

    Scheduled upkeep for flaky tests, dependency bumps and code health. Sweeps file improvement ideas as issues. Label the ones you want built.

    Be first through the door.

    Agent Factory is coming soon. Leave your email and we'll send you an invite when the autopilot is ready.