Sub-agent delegation

Agents can call other agents. The delegation graph across the workspace is visualized in the Office mode.

Two ways to delegate

Assistants — implicit delegation

An Assistant has a subagents.allowAgents list — the agents it's permitted to delegate to. The LLM decides at runtime, like any other tool call.

You are an orchestrator. When asked about code, delegate to
"Code Reviewer". When asked about customers, delegate to
"Support Agent". For anything else, answer directly.

The orchestrator pattern is the canonical use case. Give it a thin system prompt focused on routing, no tools of its own, and a short list of sub-agents.

Agents — explicit delegation

An Agent's pipeline has a Delegate node. You pick the target agent and the runtime calls it at exactly that point in the graph.

Use this when:

• The sub-agent must be called at a specific step (after data fetch, before final response, inside a loop).
• The routing decision is deterministic (e.g. based on a condition node) rather than LLM-driven.
• You need the output bound into a downstream node.

Setting up sub-agents

On the agent's Tools tab, scroll to Sub-agents and pick the agents it's allowed to delegate to. The list filters to agents in the same org.

This writes to subagents.allowAgents — a list of agent IDs. Both Assistants and Agents read from the same field; the difference is just how it's invoked at runtime.

What gets passed

When agent A delegates to agent B:

• B receives the delegation message (either a synthesized user message from A, or a structured payload from the Delegate node).
• B runs to completion with its own tools, model, and system prompt.
• B's final response is returned to A.

The proxy currently runs delegations synchronously inline — A waits for B to finish before continuing. The tool-call rounds (max 5) apply across the full call tree.

Loops and safety

Delegation cycles are technically possible (A delegates to B which delegates back to A). The runtime depth-limits delegation to prevent infinite recursion. The Office canvas helps you spot cycles visually before they bite — look for back-edges in the delegation graph.

Where to next

• Office mode
• Pipelines
• Agents and Assistants