Agents

Agent Orchestration Configuration

Updated Jun 8, 2026 · 4 min read

Agent Orchestration Configuration

The Agent Orchestration ability lets an agent spawn purpose-built helper agents. This article covers every configuration setting and how they interact.

Per-Agent Settings

These settings appear in the agent configuration panel under the Agent Orchestration ability card. Each agent can have its own limits.

Max Concurrent Spawns (default: 3)

How many forked spawns can execute at once. When the orchestrator fires off many helpers in parallel, only this many actually run their full agent loop simultaneously — the rest queue up and start as slots free. Higher values mean faster fan-out but more server load on the host machine.

Important: Blocking spawns (wait=True) bypass this limit entirely — they run inline under the caller and don't count against the concurrency cap.

Max Spawns per Agent (default: 5)

Total number of helper agents this orchestrator can spawn across a single chat session. Once the limit is hit, calls are refused. This prevents an orchestrator from accidentally or deliberately flooding the system with workers.

Counts ALL spawns created in the session — including those that finished, errored, or were stopped. Stopped/finished spawns still count against the cap.

Max Spawns per Spawn (default: 1)

How many sub-helpers a spawned clone is allowed to create. This governs recursion breadth at each level of the tree.

Max Spawn Depth (default: 2)

How many levels of spawns-within-spawns are allowed. The main agent is level 0; each spawn increments depth by 1.

Auto-Stop on Session End (default: On)

When the orchestrator's chat session ends (user closes the tab or navigates away), automatically stop all running and queued spawns. Turn this off only if you need long-running spawns to outlive the chat session — but be aware they'll consume resources with nobody to receive their results.

Default Spawn Mode (default: Fork)

How spawns run when the orchestrator doesn't explicitly pick:

The orchestrator can always override this per-spawn with the parameter.

Not Yet Configurable per Agent

These limits exist in the code but aren't yet exposed as per-agent settings — they use fixed defaults:

Clone Max Turn Count (fixed: 20 turns)

A separate turn limit applied to every spawned clone. Prevents a misdirected clone from looping endlessly on a narrow task. This is separate from the orchestrator's own on the Limits table. Twenty turns is generous for focused tasks while still catching runaway loops.

Clone Wall Clock (fixed: 300 seconds / 5 minutes)

A time cap applied to every spawned clone. Cuts off a spawn that hasn't finished within the window — catches hanging tool calls, slow models, and zombies. Separate from the orchestrator's own .

How Settings Interact

A typical tree under defaults (spawns-per-agent=5, spawns-per-spawn=1, depth=2, concurrent=3):

With spawns-per-spawn=0 or depth=1, the tree stops at level 1 — only 5 spawns total, no grandchildren.

With spawns-per-agent=3 and spawns-per-spawn=2, the tree can go deeper but narrower — 3 level-1 spawns each create 2 sub-spawns for 9 total.

Safety Notes

#agent-orchestration#configuration#spawns#limits