Pipelines
Pipelines
Pipelines carry one piece of work from a sentence to a result, in seven stages, with a decision from a person between each stage (module pipelines). An agent does each stage. You read what it did and say whether it goes on.
In the sidebar under Automation → Pipelines.
What you see
The page lists pipelines, newest first. Each card shows:
- the title, taken from the first line of the request;
- its status:
running,blocked(waiting at a gate),done,failedorabandoned; - the seven stages in order, each with where it is: not started, working, waiting for you, done in a number of steps, skipped, or the error it failed with;
- the attempt number on any stage that was sent back.
Click a title to open it: the current stage, its run trace (replayed, then followed live), the reason it was sent back, and what the stage produced. With nothing started, the page says "Nothing is being carried through yet".
The stages
| Stage | What it is for | May change things |
|---|---|---|
| plan | Work out what to do, and say what done looks like. | No |
| test | Write the checks first. | Yes |
| implement | Make the checks pass, and nothing more. | Yes |
| review | Read what was written as somebody who did not write it. | No |
| verify | Run it. Report what actually happened. | Yes |
| remember | Write down what would have helped at the start. | Yes |
| improve | Say what to do differently next time. | No |
A "No" stage runs with read-only tools, enforced. Every stage gets the full request word for word, plus what earlier stages produced. On a pipeline tied to a project, the remember stage writes facts and decisions into the project's Brain.
What you can do
Start a piece of work
- Open Automation → Pipelines and press Start work.
- Under What needs doing, write the request in full.
- Optionally pick the Project.
- Press Start. The plan stage begins.
The agent that works the stages is the one bound to the project, or to the pipeline surface. An API key cannot pick the agent.
Decide at a gate
When a stage finishes it waits for you as a request in Approvals, such as "review: Add a status page". Answer it on Approvals or from a linked chat.
- Approve marks the stage done and starts the next. Approving the last stage finishes the pipeline.
- Deny, with a reason, sends the stage back. It runs again with your reason as its new instruction. The pipeline does not fail.
A gate stays open for a week.
Follow a pipeline from elsewhere
A pipeline started from a chat reports each stage back to that chat. One tied to a task updates the task in Tasks.
Nudge or stop a pipeline
The page has no buttons for these; an agent or the API does them. Advance runs the next pending stage of a stuck pipeline and never passes a gate. Abandon stops it; stages not done are marked skipped.
Who can do what
| Permission | Label | Lets you |
|---|---|---|
pipelines:read | View pipelines and their stages | See the page, every pipeline, its stages and their traces. |
pipelines:write | Start and steer pipelines | Start work, advance a stuck pipeline and abandon one. |
Deciding a gate needs approvals:decide, as any approval does.
What happens on its own
Nothing runs on a schedule. A pipeline moves when a stage finishes and when a person answers its gate.
Needs
Projects switched on. Agents runs the stages and Approvals holds the gates; without either, a stage fails and says why.
What agents can do
No tool decides a gate: an agent that could approve its own stage would have no gate. Read-only tools run at once; the rest ask first under the default rules.
- Read-only:
pipelines.list,pipelines.get,pipelines.stages - Change:
pipelines.start(kindchangefor work on something that exists,launchfor a project from nothing),pipelines.advance,pipelines.abandon(stops it)
Over MCP the tools are named pipelines__start and so on.