Agent Control Steps
Control steps add logic and human decision points to agents. Use them to pause execution for approval, branch based on content analysis, or route work differently depending on what the agent finds.
Types of Control Steps
Agents support two control step types:
Step Type | Purpose |
|---|---|
Human input | Pause execution until a user provides input or approval |
Condition | Branch execution based on a yes/no/no-answer decision |
Control steps are inserted into the agent sequence like any other step. When the run reaches a control step, the executor evaluates it before continuing.
Human Input Steps
Human input steps pause an agent until a designated user provides input. They create approval gates, review checkpoints, or decision points that require human judgment.
How Human Input Works
When an agent run reaches a human input step:
Execution pauses: The run stops at that step and persists its current state.
Status shows awaiting input: The run viewer displays the step as "awaiting input" and indicates what input is needed.
User provides response: The designated user opens the run, reviews the context, and submits their input.
Run resumes: The agent continues from the checkpoint with the user's input incorporated into downstream steps.
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.
Configuring Human Input Steps
When authoring an agent, add a human input step to specify:
Who can provide input: Assign the step to a specific role or user.
What input is needed: Define the question or decision required—such as approval, selection from options, or free-form text.
Context shown: Specify what documents or run outputs the user should see when reviewing.
Human input steps require someone to actively respond. Agents with pending input steps remain paused until the assigned user acts or the run is cancelled.
Condition Steps
Condition steps branch agent execution based on an evaluated decision. The condition typically evaluates document content, run state, or analysis results to determine which path to take.
How Conditions Work
Conditions produce a ternary result:
Yes: Follow the "yes" branch.
No: Follow the "no" branch.
No answer: The condition couldn't be determined—follow a fallback path or flag for review.
The agent evaluates the condition and routes to the appropriate branch without pausing.
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.
Configuring Condition Steps
When authoring an agent, add a condition step to specify:
The condition to evaluate: Define what the agent should check—typically based on document content, extraction results, or run metadata.
Branch behavior: Define what happens for yes, no, and no-answer outcomes.
Combining Control Steps
You can use human input and condition steps together to create sophisticated agents:
Use a condition step to detect whether a document requires special handling.
If yes, route to a human input step for specialist review.
If no, proceed directly to output generation.
This pattern lets you automate the straightforward cases while escalating exceptions to humans.
Paused Runs
Only human input steps pause agent execution. Condition steps evaluate instantly and branch without stopping.
When a run is paused at a human input step:
The run appears in your run list with status "awaiting input."
Open the run to see what input is needed.
Provide your response and submit.
The run resumes automatically and continues through remaining steps.
Run History and Auditing
Control steps are logged in run history. Each run record shows:
Which control steps were encountered.
What condition was evaluated and what result it produced.
Who provided input for human input steps and when.
How long the run spent paused at each checkpoint.
This creates an audit trail for agents that involve human decisions.
Best Practices
Use human input for true decisions: Add human input steps where human judgment genuinely adds value—not for steps that could be automated.
Keep conditions clear: Write conditions that produce unambiguous yes/no results. If the outcome is often "no answer," reconsider how you're evaluating.
Assign input steps thoughtfully: Route human input to the right person—partner review for strategic decisions, associate verification for quality checks.
Test branching logic: Run test inputs through condition steps to verify each branch produces the expected outcome.
Document branch behavior: In your agent description, note what happens under each condition outcome so team members understand the logic.
What to Read Next
Agents Overview for building and running agents
Agent Triggers for scheduling and event-based automation
Playbooks to apply firm standards in agent steps