No written spec
So the human keeps filling it in with corrections.
An Automattic project
Stop babysitting agents. Run pipelines of them.
One task, six phases, a different team of agents in each. Every phase writes a file you can read, fix, and re-run from.
- 6phases
- 15agents shipped
- 2CLIs supported
The problem
Most of us still sit next to the agent.
Steering an agent in real time works, but it caps how much one person can ship.
Non-deterministic
Same prompt, different run, different result. You stay in the loop just in case.
No structure
Tests and docs happen only when someone remembers to ask.
Invisible to the team
Context lives on one laptop. Everyone else sees the PR at the end.
What we're building
An orchestrator that runs the work through phases.
Six phases, an artifact at each, agreed up front. Runs on its own. Stop, fix any phase, re-run from there.
Runs without you
Decisions agreed at the start. Executes to the target phase without checking back in.
Writes everything down
Each phase commits its file before the next starts. Your team reads it as it lands.
Restartable from any phase
Bad spec? Fix the spec. Bad plan? Fix the plan. Re-run from that point, not from zero.
More tokens where it matters
Adversarial reviewers, parallel writers, cross-provider runs. Spend compute on the risky surfaces.
What it actually does
Files in your repo, one phase at a time.
No new UI. The orchestrator runs in your CLI (Claude Code or Pi) and writes artifacts into a pipeline folder in the repo. Your team reads them like any other file.
terminal · claude-code
demo · sped up
.pipelines/issue-1234/base/
Reconstructed log of a real pipeline run, sped up. Every phase pairs
a writer with an adversarial reviewer; both commit their artifacts
(spec.md, spec-review-approved.md, …) so
the team can see what got pushed back on and why.
Run it yourself →
Why it matters
The bottleneck moves from supervising to reviewing.
More work in flight
One person runs many pipelines at once. Review when done.
Fixes that stick
Patch the phase that failed. Every future run benefits.
Tests and docs by default
Built into the pipeline. Stop relying on whoever's driving.
Visible before the PR
Each phase's artifact lands in the repo. Team can read along.
Why now
The pieces are in place.
Agents are good enough.
Given a clear spec, they do the job. The gap is process, not model.
Human time is the bottleneck.
Every hour assisting an agent is an hour not spent on work that needs a person.
Tooling caught up.
Claude Code and Pi already ship skills, teams, hooks, worktrees. We build on top.
Try it
Install in one line.
Claude Code
/plugin marketplace add Automattic/radical-pipelines
/plugin install radical-pipelines@automatticPi
pi install git:github.com/Automattic/radical-pipelines
Pi also needs the phase agents available in .pi/agents/ or
~/.pi/agent/agents/.
See README for full setup →
Built at Automattic. Open source.
Code, agents, skill, methodology, all in one repo.