OpenClaw swarm orchestration is the system that lets one OpenClaw agent spawn and coordinate multiple subagents working concurrently — and as of the OpenClaw 2026.9.1 release, published 3 September 2026, it is switched on by default. That single change matters more than it sounds: parallel multi-agent work went from an opt-in power feature to the standard behaviour of every up-to-date OpenClaw install. This guide covers how the swarm machinery actually works, the spawn and collection tools involved, the concurrency limits that keep it from running away with your machine, and the config keys to tune — all drawn from OpenClaw's official release notes and subagent documentation.
📺 Watch: OpenClaw 2.0 Is a MASSIVE Upgrade for AI Agents
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside · Want AI SEO help 1-on-1? Book a free SEO strategy session →
OpenClaw Swarm Orchestration: What Changed in 2026.9.1
Per the official release notes on GitHub, OpenClaw 2026.9.1 ships "swarm orchestration enabled by default for concurrent sub-agents". Before this release, concurrent subagent work was something you configured deliberately; now a fresh or updated install coordinates parallel child agents out of the box. The same release added personal skill libraries on shared Gateways via the openclaw skills library command — covered in the OpenClaw skills library guide — plus Mermaid diagram rendering in the Control UI, configuration-preserving updates with rollback, and an Android UI that finally matches the desktop Control UI. Two days later, the 2026.9.2 release of 5 September added GPT-6 Astra support and settings changes without restarts, so the swarm layer now sits inside a noticeably more stable shell.
How OpenClaw Subagents Actually Work
Under the hood, OpenClaw's documentation describes subagents as background agent runs in isolated sessions. A parent agent spawns one with the sessions_spawn tool, which returns a run id as soon as startup is accepted — it does not block waiting for the child to finish. Each spawn takes a required task description, plus optional parameters: which agent to run it as, a context mode of isolated (the default) or fork (which branches the parent's transcript into the child), a visible flag that creates a persistent dashboard session for user-facing work, and per-run overrides for model, thinking level and timeout.
Results come back push-style. Rather than polling children, the parent calls sessions_yield to end its current turn and wait for announced completion events, which arrive as its next message. The completion handoff includes the child's final assistant text, its status — completed, failed or timed out — and runtime and token stats. For inspection while things run, the docs list slash commands: /subagents list to see current runs, /subagents log to read a child's turns, and /subagents info for metadata.
If you want multi-agent systems that actually make money instead of just burning tokens in parallel — check out the AI Profit Boardroom → get the Agent OS and swarm playbooks. Prefer a 1-on-1 route? Book a free SEO strategy session and plan your build.
Swarm Collection vs Announcements
One distinction in the documentation is worth getting right, because it changes how you structure orchestration code. Ordinary subagents announce their results back to the requester automatically. Swarm collectors work differently: per the docs, swarm results come back "through explicit collection, not completion notifications" — the orchestrator gathers children with agents_wait or an awaited agents.run call. In practice that means announcements suit fire-and-forget delegation where the parent reacts as results land, while explicit swarm collection suits batch jobs where you fan out a set of tasks and want all the results together before the next step. Choosing the wrong pattern is how people end up with orchestrators that stall waiting for events that were never going to arrive.
The Limits That Keep a Swarm Under Control
Default-on concurrency without guardrails would be chaos, and OpenClaw's documentation is unusually specific about the guardrails. The key config family is agents.defaults.subagents, with per-agent overrides available under agents.entries. The documented defaults: maxSpawnDepth is 5 (children can spawn grandchildren, but only five levels deep), maxChildrenPerAgent is 5 active children per agent session (raisable to 20) to prevent what the docs call "runaway fan-out from a single orchestrator", and maxConcurrent is 8 across the dedicated in-process subagent queue lane. Announce timeouts default to 120,000 milliseconds, and finished child sessions are archived after 60 minutes.
There is also backpressure on the delivery side: OpenClaw warns when the result-delivery backlog reaches 25 and blocks new subagent spawns entirely at 50 until an operator retries or dismisses enough retained deliveries. If your swarm suddenly refuses to spawn children, the backlog cap is the first thing to check — it is a protection mechanism, not a bug.
📺 Watch: NEW OpenClaw 2.0 Destroys Hermes Agent?
What Swarm Orchestration Is Actually Good For
The pattern shines wherever work is naturally parallel and each unit is independent: researching a list of prospects, auditing every page of a site section, drafting variations, or monitoring several sources at once. It is the same shift this site covered when Hermes agent swarms arrived, and the economics logic mirrors the Kimi K2.6 agent swarms analysis: parallelism buys wall-clock time, not intelligence, so a swarm of mediocre agents just produces mediocre output faster. The per-child model override is your lever here — cheap models for the fan-out work, a stronger brain for the orchestrator that judges the results.
Model choice for those seats is its own decision, and the Goldie Bench write-up covers how the current agent brains compare in hands-on tests. It also pays to remember that scheduled and recurring jobs interact with swarms — a cron-triggered orchestrator spawning children on a timer is a powerful but easy-to-overdo pattern, and the agent cron job guide covers the scheduling half of that equation.
Running Your First Swarm Without Regretting It
A sensible first swarm uses the documented parameters conservatively. Start with the context choice: isolated is the default and the right call for independent tasks, because each child begins clean and cannot be confused by the parent's history; fork copies the parent's transcript into the child, which suits work that genuinely needs the conversation so far — a follow-up analysis, say — at the cost of a bigger context. Set visible to true while you are learning, since persistent dashboard sessions let you watch children work instead of guessing. Give every spawn a runTimeoutSeconds you have actually thought about, because a hung child otherwise waits out the defaults, and remember the announce timeout is two minutes — an orchestrator that yields for children doing twenty-minute jobs needs its expectations configured, not assumed. Finally, resist raising maxChildrenPerAgent until a real workload has hit the ceiling of five: the documented caps exist because fan-out multiplies token spend as efficiently as it multiplies work, and the docs' phrase "runaway fan-out" describes a bill as much as a bug. Archived children disappear from view after 60 minutes by default, so pull anything you need from /subagents log before it goes.
Swarms Across the Wider Agent Landscape
OpenClaw defaulting swarms on is part of a broader 2026 trend: every serious agent platform has been converging on multi-agent operation as the normal mode. Hermes v0.21.0 shipped Bot Mode's society of named agents with group chats at the end of August, and its release notes describe live steering of subagents mid-flight. The full v0.21 breakdown covers that side. The practical takeaway is the same on every platform: the bottleneck has moved from "can I run agents in parallel" to "can I specify tasks well enough that parallel agents produce work worth collecting". Orchestration is now free; judgement is not.
Frequently Asked Questions
Do I need to enable swarm orchestration in OpenClaw?
Not any more. Per the 2026.9.1 release notes of 3 September 2026, swarm orchestration for concurrent subagents is enabled by default. On older installs, updating brings the new default — and the same release made openclaw update preserve settings and roll back failed candidates.
How many subagents can OpenClaw run at once?
The documented defaults are 8 concurrent subagents in the queue lane, 5 active children per agent session, and a spawn depth of 5, all configurable through the agents.defaults.subagents keys — plus a hard block on new spawns when the delivery backlog hits 50.
Can each subagent use a different model?
Yes — sessions_spawn accepts a per-run model override, and defaults live at agents.defaults.subagents.model. Cheap brains for fan-out work, a strong brain for orchestration is the standard economics.
The Bottom Line
OpenClaw swarm orchestration going default-on in 2026.9.1 makes concurrent multi-agent work the baseline, with sensible documented limits — depth 5, 5 children per agent, 8 concurrent — and explicit collection for batch fan-outs. Learn the spawn, yield and collect patterns, tune the caps to your machine, and put the parallelism to work on tasks that pay.
If you want swarm workflows that produce revenue — prospecting, content, audits — already built inside Agent OS, check out the AI Profit Boardroom → join the community building this daily. Want tailored advice first? Book a free SEO strategy session.
