Agents

Agent Control Steps

Control steps add human decision points and branching logic to an agent. A Question step pauses the run until a person replies, and a Branch step sends the run down one of several paths based on a rule you describe.

Types of Control Steps

Step Type

Purpose

Question

Pause the run and ask a person for input or approval

Branch

Choose which path the run takes next, based on a rule

You add them like any other step, either by asking for them when you create or edit the agent in chat, or with Edit manually on the canvas. See Build and edit agents on the canvas.

Question Steps

A Question step stops the run until someone answers, which creates an approval gate or review checkpoint where human judgment matters.

Set up a Question step

On the canvas, set the step type to Question and fill in What to ask the person, for example "Approve this draft to continue?". The question is the only setting: anyone who can see the run can answer it, and the reply is free text.

What happens during a run

  1. The run reaches the Question step, saves its progress, and shows as Waiting.

  2. The run page shows the question with a reply box. If the run belongs to a chat, the question also appears there.

  3. Someone replies, and the run continues with the answer available to later steps.

How to reply, who can reply, and how long a run can wait are covered in Resume an agent run waiting on human input. Waiting runs also affect when the agent's scheduled runs start; see Agent Triggers.

Common Use Cases

  • Partner review before delivery: Draft the document, pause for partner sign-off, then finalize and deliver.

  • Client approval gates: Generate options, pause for client selection, then proceed with the chosen approach.

  • Quality checkpoints: After AI extraction, pause for an associate to verify key fields before generating final outputs.

  • Escalation decisions: If AI flags ambiguous terms, pause for a lawyer to clarify before proceeding.

Branch Steps

A Branch step evaluates a rule and continues down the matching path without pausing.

Set up a Branch step

On the canvas, set the step type to Branch and fill in:

  • How should it decide? The rule to evaluate, for example "Is the contract value over $1M?".

  • Paths: a new Branch step starts with Yes and No. Click Add path for more, such as one path per issue type. For each path, set a name, When the result is (the result that selects it), and Go to (the step where the path begins).

  • Path rejoin behavior: whether the paths merge back into one flow afterwards. See Build and edit agents on the canvas.

Steps on paths the run did not take are skipped and appear faded on the run graph.

Common Use Cases

  • Clause routing: If the document contains a change-of-control clause, route to the M&A specialist review path. If not, proceed with standard review.

  • Threshold decisions: If the extracted liability amount exceeds a threshold, flag for partner review. Otherwise, proceed to summary generation.

  • Issue classification: Based on the issue type found, route the document to the appropriate work stream.

  • Quality gating: If AI confidence is low on an extraction, flag for human verification. If high, proceed to output.

When paths end in different deliverables, give each path its own output. See Agent output contracts and branch-specific outputs.

Combining Control Steps

Use both step types together to automate the straightforward cases and escalate the exceptions:

  • Use a Branch step to detect whether a document requires special handling.

  • On the Yes path, add a Question step for specialist review.

  • On the No path, proceed directly to output generation.

Best Practices

  • Ask only for real decisions: Add Question steps where human judgment genuinely adds value, not for steps that could be automated.

  • Write clear rules: Phrase each Branch rule so its result maps cleanly onto one path.

  • Ask specific questions: Say exactly what the reply should contain, such as "approve" or a counterparty name, so whoever answers knows what to write.

  • Test every path: Run test inputs that exercise each path to verify it produces the expected outcome.

  • Document branch behavior: In your agent description, note what happens on each path so team members understand the logic.

Was this helpful?
Was this helpful?