Alphav0.1 · 157e418

auto is a programmable software factory — defined in YAML, woken by events, steered by you, improved by itself.

01 / Define

Your org chart is a directory.

auto is a version-controlled workforce: agents, runtimes, triggers, and policy live beside your application in .auto/. Merge a pull request and the factory updates — just like your CI/CD.

.auto/

agents/

fragments/

agents/chief-of-staff.yaml

An orchestrator that lives in Slack — tag it with work and it runs the fleet.

1name: chief-of-staff
2imports:
3  - ../fragments/environments/agent-runtime.yaml
4session:
5  archiveAfterInactive:
6    seconds: 86400
7systemPrompt: |
8  You are the Chief of Staff Engineers: a
9  one-live-session orchestrator that lives in
10  Slack. Humans tag you with lists of tasks.
11  Dispatch one staff-engineer session per
12  task, shepherd every session until its PR
13  has green CI and a clean pr-review verdict,
14  and deliver one collated packet back.

02 / Trigger

The factory doesn't wait to be prompted.

A pull request opens. A check fails. An incident pages. A cost metric spikes. auto routes each event to the agent that owns the work — starting a session when there is none, or rejoining the one that already exists.

chief-of-staffsession started

triagesession started

designersession started

staff-engineersession started

incident-responsesession started

pr-reviewsession started

self-improvementsession started

03 / Operate

The work keeps one memory — even when it moves.

Steer from the dashboard, Slack, or the PR. The agent carries the same context everywhere; whoever on your team jumps in, nothing gets re-explained. When a decision needs a human, your agents ping you.

staff-engineersession · running

04 / Evolve

The factory files PRs against itself.

The factory is all software. Agents study their own sessions, improve their configuration, and author new automations when the line needs more capacity. Every change ships as a pull request; you approve with a merge.

.auto/agents/pr-review.yamlself-pr opened by introspector · +5 −2

  @@ learned from 47 review sessions @@  systemPrompt: |    Review pull requests like a senior engineer.+   Trace every changed callsite before approving —+   the last two regressions shipped through+   untraced call graphs.− model: claude-haiku-4-5+ model: claude-fable-5  triggers:    - events: [github.pull_request.opened]−     reviewDepth: quick+     reviewDepth: exhaustive

Reviewed like any other change✓ your merge is the approval

Cloud agents that hold up on a real codebase.

Your setup, sandboxed

Declare your environment in .auto/ like a Dockerfile — Docker-in-Docker included. Every session spins up fresh, runs isolated, and tears down clean.

Durable sessions

Every unit of work is a session — live transcript, observable state, resumable long after it ran.

Secrets, sealed

Envelope-encrypted, injected at runtime — or at the network boundary by a proxy, so your agent never holds the key at all.

Any harness, any model

Each agent is Claude Code or Codex — the same harness you run in your terminal — with 300+ models to swap between on a dime, open source or frontier.

Install a crew off the shelf — or build your own.

Start with the same teams of agents we use internally, and remix them to your liking. Or start from scratch and tell your agent your ideal automation suite — they'll build the right .auto/ directory for you in a pull request.

Product work

The Accelerator

Turns a goal into owned, review-ready pull requests — and keeps the queue moving.

Incident response

The War Room

Triages alerts, investigates causes, and drives fixes all the way through resolution.

Code health

The Slopbusters

Continuously simplifies, repairs, and documents the codebase as it changes.

Chief of Staff Engineers

Takes a task list, dispatches staff engineers, shepherds them to green.

Your tools are their tools.

Agents use your tools and listen for your events. Hook into anything with an MCP API or an event stream — or stand up your own when you need to.

GitHub

Agents open PRs, answer review comments, and turn red checks green.

Slack

Agents take work from a mention and report back in the thread.

Linear

Agents triage incoming issues and carry tickets through to done.

…and any MCP server, any webhook

Recently shipped

Full changelog →

Jul 20

More control over access and funding

Project settings show each OpenAI subscription quota window and reset time separately, so available usage is easier to understand.

Jul 16

Faster navigation and clearer recovery

Onboarding recognizes when an organization is already set up, so returning teams go straight to their sessions instead of repeating the welcome flow.

Jul 14

Settings rebuilt around your organization

Billing and GitHub configuration now live in one consolidated settings area instead of being scattered across the app.

Coding agents made the coding fast. The rest of the lifecycle still bottlenecks on people: writing prompts, routing work, chasing follow-ups, reviewing diffs.

The best engineers program away these bottlenecks — they build the factory that builds the product.

auto gives you the tools to do that so you can apply your judgement where it matters most — and let agents handle the rest.

auto

Your next teammate is a pull request away.

Request access for a factory of your own. Invites go out in small waves while the alpha hardens.

Alpha — built for agents, by agents