A pipeline is a repeatable, multi-step workflow that chains agents, scripts, quality gates, and reports into one process you can run again and again. Where a single chat does one thing at a time, a pipeline runs a whole sequence — for example, plan → implement → review → fix — with as much or as little configuration as you want. This article explains what pipelines are, how they run, and where to build them.
Where pipelines live
The Pipelines page is both a library and an operations console: it holds your pipelines and shows their runs. From the library you can create pipelines from YAML, duplicate and edit them, enable or disable which ones show up in the run pickers, and share pipelines with your team through the cloud.
How a pipeline runs
Pipelines run in one of three modes:
- Dispatch — you start it manually.
- Plan-ready — it runs from a plan you've already made, skipping the planning step.
- Task-ready — you hand it a task and it writes the plan itself first.
Pipelines can also fire automatically on events — CI, build, and source-control events — so a run kicks off when, say, a check fails or a pull request is opened. Once running, a pipeline works in the background: you can have several going at once and check on each from the execution list.
Build pipelines with the Pipeline Assistant
You no longer have to hand-write pipeline YAML from scratch — the Pipeline Assistant is now available. It's a dedicated chat on the Pipelines page: open a new chat, describe the pipeline you want in plain language, and it helps you build and refine it. The Pipeline Assistant is separate from your worktree chats (which run coding tasks) and from Cowork chats — it's scoped to building and debugging pipelines. You can still edit the YAML directly whenever you want.
What pipelines are good for
- Deep, repeatable workflows — for example, a full implementation pipeline that plans, partitions, implements, and QA-loops a feature.
- Multi-model code review — run reviews, including security checks, across agents and post the findings as pull-request comments.
- Keeping work moving — fire off a pipeline and let it run in the background while you do something else.
Next steps
- Anatomy of an implementation pipeline — the full lifecycle of a deep pipeline, stage by stage.
- Reading a pipeline execution — follow a run and act on the result.
- Hooks: automate the busywork — fire pipelines and actions automatically on events.