Sequences — how your AI Team gets work done

An AI Team is your agents and the work you've given them. A Sequence is the order that work happens in, and what happens when things differ.

Every project has one Sequence. Until you add a branch it's a straight line: step one, step two, step three, each able to see what the ones before it produced.

What starts it is a trigger — a clock, or the client's own system calling in. Without one, a Sequence only runs when you press Run.

The four kinds of step

You choose, step by step, whether you want judgement or certainty.

Step	What it is	Uses a model?
Task	An agent is given the job and its tools, and works out how	Yes
Action	One tool, arguments you set, run exactly as written	No
Condition	Splits the Sequence — yes one way, no the other	Only if you ask it to
Agents Route	Sends work to whichever specialist should handle it	Only if you ask it to

An agent can also be a step in its own right — work handed straight to it, with its role and goal as the instruction, rather than a task written out. That's what makes a supervisor sending work to specialists worth drawing.

Add them from the canvas toolbar. Each has its own article above.

Where a run begins

With no trigger, at the top. With one trigger, wherever that trigger points.

With several triggers pointing at different steps, each one enters where you drew it: a nightly sweep starts in one place, a customer's cart posting in starts in another. That is never guessed — the thing that fired says where it goes.

Wiring the Sequence

No branches? Steps run top to bottom, in order. Nothing to wire.
Once anything branches, every step says where to go next. A step with nowhere to go ends the run — that's deliberate, because the step sitting after a "yes" branch is usually the "no" branch, and guessing would run both sides of your condition.

You can point a step back at an earlier one to make a loop — a reviewer that sends work back to be redone. Loops are capped at 40 steps, so a review that never approves stops rather than running all night.

Joining steps up is done on the canvas by dragging. See Working on the canvas.

When something goes wrong

A Sequence tries to finish rather than stop dead:

A tool fails (bad reference, service down) — the failure is recorded as that step's result and the run continues. You'll see it in the trace.
A tool isn't connected — the step says so and points you at Tools to add the key.
A step has no agent — it doesn't run, and says so. Nothing is guessed.
A condition can't work out an answer — it takes No, the branch that doesn't send anything.
A route matches nothing — it takes Otherwise, which always leads somewhere you chose.

A worked example

Workspace booking support:

Task — Front Desk agent, with the payment, knowledge and WhatsApp tools: "Find out what happened with this customer's booking."
Condition — Compare text: does {{previous}} contain "success"?
If yes → Action — send the booking code by WhatsApp.
If no → Task — Front Desk: "Draft an apology and explain how to rebook", grounded in the refund policy from the knowledge base.
Then → Action — send {{previous}} by WhatsApp.

The decision is on the canvas where you can see it and change it, rather than buried in an instruction and hoped for.

For a support desk that has to reach three different specialists rather than two outcomes, replace step 2 with an Agents Route.

Questions people ask

How many Sequences can a project have? One. If you need two genuinely separate flows, make them two projects.

Can two branches join back together? Yes — point both at the same step.

How do I pass one step's result to another? Tasks read the work so far automatically. Actions use {{previous}} and friends — see Action.

Where do I see what actually happened? The Execution tab, step by step, including why each decision went the way it did. See Running agents & reading traces.