Sub-Agents¶

A sub-agent is a workflow step that spawns a child workflow to handle a focused sub-task. Think of it as delegation: the parent workflow describes an objective in natural language, and the sub-agent plans and executes its own steps to accomplish it, then returns the result.

This is useful when part of a workflow requires deeper reasoning, a separate set of tools, or a different perspective than the main flow.

[Screenshot needed: Agent step in the workflow editor showing objective, briefing fields, and the parallel execution tab]

When to Use Sub-Agents¶

Sub-agents suit tasks that are too complex for a single step but have a clear deliverable:

Research: "Summarise the top 5 competitors in this market"
Analysis: "Review this contract and flag any unusual clauses"
Content generation: "Draft three LinkedIn posts from different angles on this topic"
Data processing: "Collect pricing data from these three sources and produce a comparison table"

For simple operations (formatting a value, filtering a list, calling a single API), a code step or integration step is more appropriate.

Configuring a Sub-Agent Step¶

In the workflow editor, an Agent step has the following fields:

Objective: A natural language description of what the sub-agent should accomplish. This supports variable interpolation from previous steps, so you can pass dynamic context.
Briefing (optional): Key facts, constraints, and role information that shape the sub-agent's behaviour. For example, you might specify "You are a financial analyst" or "Only use data from the last 30 days".
Allowed integrations (optional): A whitelist of services the sub-agent may use (e.g., "slack", "gmail"). If omitted, the sub-agent can use all services available to you.
Budget limit: The maximum the sub-agent may spend. Defaults to $1.00, with a range of $0 to $100. The parent workflow's remaining budget is divided among its children.

Parallel Execution¶

Sub-agents can run multiple child workflows at the same time. There are two modes for this.

Split work across children¶

Set a parallel count (up to 5) and optionally a variable to vary across the children. For example:

Objective: "Research {{competitor}} and summarise their pricing strategy"
Variable: competitor
Values: ["Apple", "Google", "Microsoft"]

This spawns three child workflows, each researching a different competitor. The budget is split evenly among them.

Multi-perspective mode¶

Define named perspectives, each with a role, area of expertise, and focus. For example:

Legal Analyst: focus on compliance risks
Financial Analyst: focus on cost implications
Technical Lead: focus on integration feasibility

Each perspective runs as a separate child workflow, and the results are aggregated. You choose how results are combined:

List: Returns an array of all outputs
Merge: Concatenates the text outputs
Best: Returns the first successful result

Budget and Safety Limits¶

Sub-agents operate within several guardrails:

Limit	Value	Purpose
Default budget per agent	$1.00	Prevents runaway spending on a single sub-task
Budget range	$0 -- $100	Configurable per step
Maximum nesting depth	10	Prevents unbounded recursive spawning
Inactivity timeout	10 minutes	Terminates sub-agents that appear stuck
Maximum identical retries	3	Stops repeating the same failing action
Depth warning threshold	5	Logs a warning (does not block execution)

Budget is the primary constraint. It naturally limits how much work a sub-agent can do. The depth limit and inactivity timeout are infrastructure safeguards that protect against edge cases.

When a sub-agent exhausts its budget, it stops and returns whatever partial results it has produced. The parent workflow continues with those partial results.

Enabling and Disabling Sub-Agents¶

You can control whether your workflows are allowed to spawn sub-agents through the collaboration.allow_agent_spawning toggle in Settings under Collaboration → Allow Dynamic Agent Spawning. This is disabled by default. You must enable it before using Agent steps. If it is disabled, any Agent step in a workflow will return an error explaining that dynamic agent spawning is turned off.

Sub-Agents in Virtual Run¶

During a Virtual Run preview, Agent steps are simulated. You see what the sub-agent would produce, including:

The generated child workflow plan
The objective and briefing sent to the child
The expected cost breakdown

This lets you verify that the sub-agent's scope and budget are appropriate before running the workflow live.