Task — when you want judgement

A task hands a job to an agent and lets it work out how. The agent decides which of its tools to use, in what order, and when it has enough.

Add one from the canvas toolbar: Add task.

Give it the goal, not the steps

"A customer says they paid but got no booking code. Find out what actually happened."

The agent checks the payment, notices it was abandoned rather than failed, checks their history, sees three previous successful bookings, and writes that up. You didn't script any of that, and you couldn't have — you didn't know the payment was abandoned when you wrote the task.

Use a task when the right move depends on what it finds. When there is nothing to decide, use an action instead.

Every task needs an agent

Join an agent to the task by dragging from the agent's right side to the task's left side, or pick one in the task editor.

A task with nobody on it shows Pick an agent in red, and it will not run. The step stops, the trace says why, and the rest of the Sequence carries on without it.

That is deliberate. It used to be handed quietly to whichever agent happened to be first in the project, which is how a customer's escalation ends up written by your invoicing agent, with nothing anywhere saying so.

What the agent can see

Each step can see what the ones before it produced. You don't have to pass anything along by hand for a task — the agent reads the work so far as context.

Expected output

Optional, and useful when the next step needs a particular shape: "A short paragraph, no greeting", "Just the booking reference". It's guidance to the agent, not a rule the system enforces.

Questions people ask

Are an agent's roles fixed? No — you type the role yourself. There's no list to pick from: Customer Support Specialist, Debt Recovery Officer, Front Desk, whatever fits the job.

Does an agent's role change during a run? No. An agent has one role, one goal and one set of tools from start to finish. What changes is what it has read by the time it gets to a later step.

Can one agent perform several steps? Yes, as many as you like. Join it to each of them.

What if I don't assign an agent? The step won't run and the trace will tell you which one it was. Nothing is guessed.

How do I stop it wandering off? Give it fewer tools. An agent only uses what it's been given, and a tight set of three is more reliable than a hopeful set of ten.

Related: Sequences · Action · Designing agents