⚙️ Autonomous agents · Cloudflare Workers · GitHub Actions

Describe a goal.
A team of AI agents ships it.

Factory turns a plain-English goal into a complete software delivery loop: a planner agent decomposes it into stages, developer and QA agents execute them in real GitHub Actions runs, and an evaluator agent independently verifies the result — retrying automatically until it passes.

Runs on free LLM tiers (OpenCode Zen with Groq fallback) · no servers, no VMs, no setup

What Factory does

You write the goal. Factory plans, builds, tests, verifies and reports — with a full audit trail of every stage, commit and test result.

🧠

Goal-first planning

A planner agent reads your goal and produces a minimal multi-stage plan, validated by the orchestrator before any code is touched.

🤖

Five specialized roles

Planner, developer, QA, evaluator and fixer — each a single-purpose agent with its own prompt, guardrails and expected output format.

🚀

Real CI execution

Every stage runs as a GitHub Actions job on your repository. Agents edit files, run the test suite and commit with [skip ci].

Independent verification

An evaluator agent re-checks acceptance criteria, behavior and tests, and issues the final PASS / FAIL decision — never the author.

🔁

Self-healing retries

Failures trigger a fixer → QA → evaluator loop (up to 3 attempts). Broken JSON, timeouts and failing tests are all absorbed automatically.

🛡️

Guardrails built in

Agents may only touch the work/ directory, can't traverse paths, never remove tests and never touch credentials.

🪶

Stateful orchestration

A Cloudflare Durable Object tracks each goal's state machine — stages, retries, history — across all CI runs. Nothing is lost on crash.

📊

Full observability

Every stage, commit hash, test result and agent decision is recorded and shown live in this dashboard as the run progresses.

How it works

One goal flows through four phases, end to end, with no human in the loop.

Describe the goal

Write a task in plain English — “make the tests in work/ pass”, fix a bug, add a feature. Long, detailed goals are welcome.

Planner designs the plan

The planner agent returns a minimal 1–4 stage plan (developer → … → evaluator). The orchestrator validates it before executing.

Agents execute in CI

Each stage is dispatched as a GitHub Actions run. Agents inspect the repo, write code in work/, run tests and commit.

Evaluator decides

An independent evaluator verifies the result. PASS ends the goal; FAIL starts the fixer → QA → evaluator retry loop.

Built on

Production Cloudflare primitives + GitHub CI, so the whole factory is serverless and free-tier friendly.

Cloudflare Workers — API & router Durable Objects — goal state machines Workers KV — goal index GitHub Actions — agent execution OpenCode Zen — free LLMs Groq — fallback provider Rust agent-runner — one binary, five roles

Ready to build something?