Run · Agent pipelines
Agents that hand off work to agents, never past a human who's watching.
A pipeline is several agents, each doing one part of a job well, handing off to the next when their part is done. Put an approval checkpoint on any handoff you choose, and a human stays above the loop exactly where it matters.
Why one agent isn't enough
One prompt, one voice, one blind spot.
A single all-purpose agent carries one set of guardrails and one knowledge base into every part of a job, whether it's reading an inbound request, drafting a reply, or sending it. Widen its instructions to cover every step and you widen what it's allowed to get wrong at every step, too.
A pipeline scopes each step to its own agent instead: its own guardrails, its own knowledge, and, where you decide it's worth the pause, its own human checkpoint before the work moves on.
How it runs
One request, several agents, checkpoints where you put them.
Every arrow is a handoff, logged the moment it happens. Every gate is optional: shown here to make the point, not because a real pipeline needs one on every step.
- Approval checkpoints
- Switched on per handoff, not per pipeline. Only the step where a mistake would be expensive gets held.
- Full audit trail
- Every handoff and every approval writes to the same hash-chained log as the rest of the platform.
- Governed by templates
- Each agent in the pipeline keeps its own guardrails and knowledge base. Nothing is shared by accident.
- Visible in the fleet
- Every stage is its own agent in your fleet grid, with its own status, cost and owner. Never a black box.
A checkpoint hands off to human-in-the-loop review; what a reviewer's correction teaches the rest of the company is Insight elevation.
Start on an event
A pipeline can start itself when something happens.
Not every pipeline waits for a request. Pick an event as the trigger, and the pipeline runs the moment it happens, so the right people hear about it without anyone watching a dashboard.
- A watch word fires
- Post the flagged conversation to your team's Slack channel straight away.
- A review item is created
- Notify the person who owns that queue, or the reseller who looks after the client.
- Your oversight, their AI
- A reseller can bring a client's issues to its own team without logging in to every account.
FAQ
Companies building their first pipeline are asking
How is a pipeline different from one agent with a long prompt?
One agent with a long prompt still has one set of guardrails, one knowledge base and one voice, no matter which part of the job it is doing. A pipeline is several agents, each scoped to one step with its own guardrails and knowledge, handing the work to the next one when its part is done.
Does every handoff need human approval?
No. An approval checkpoint is something you switch on for a specific handoff, typically the one step where getting it wrong is expensive. The rest of the pipeline runs on its own.
What happens while a handoff is waiting on approval?
The work sits in the reviewer's queue with the full context that produced it. Nothing moves to the next agent until someone approves it, or sends it back.
Is a pipeline harder to audit than a single agent?
The opposite. Each handoff and each approval is its own entry in the same hash-chained audit trail as the rest of the platform, so a pipeline actually leaves a more detailed record than one agent quietly doing the whole job itself.
Can a pipeline start without a request?
Yes. A pipeline can start on an event instead: a watch word firing or a new item in the review queue, for example. Its steps can notify a person or post to a Slack channel, so a flagged conversation reaches the right team the moment it happens.
Give your busiest process a pipeline, not a prompt.
Twenty minutes with the founder. No slides. The real product.
