Introducing Swarm Mode: agents that work as a team
One agent can do a lot. A coordinated team of agents can do what an entire department does. Swarm Mode spawns parallel workers under a single orchestrator — so your hardest workflows ship in hours, not weeks.
Until now, a Spawn session has been one agent, one job. That model is clean and it's carried us a long way — but there's a shape of work it can't touch: the kind where you need five people working in parallel, each on a different piece, coordinating as they go. Tender evaluations. Competitive landscape reports. Multi-vendor RFQs. Any task where a single agent, no matter how capable, becomes a bottleneck.
Today we're shipping Swarm Mode — one prompt, many agents, one coordinated result.
What Swarm Mode is
You describe a task the same way you always have. The difference is what happens next. Instead of one agent taking the task end-to-end, an orchestrator reads it, decides the right team shape, and spawns worker agents — each in its own isolated sandbox — to handle the pieces in parallel.
A market research task might spin up four workers: one per competitor. A sales account plan might spin up three: a researcher, a personalization specialist, and a writer. The orchestrator watches the team, passes updates between them, and assembles the final deliverable once they're done.
How it works
Three moving parts. That's the whole mental model.
- The orchestrator — reads your task, decomposes it along the right seams, and spawns workers. High-agency: it names the team, writes each worker's mission brief, and keeps them aligned.
- The workers — each one a full Spawn agent with the complete toolkit (artifacts, code, browsing, integrations, image/video generation) running in its own sandbox with root access.
- Group chat — a shared channel visible to every worker, the orchestrator, and you. Workers post progress. The orchestrator redirects when needed. You jump in at any point.
Behind the scenes, every worker runs on its own Modal sandbox but shares a common filesystem volume, so an agent producing intermediate files can hand them to the next worker without re-uploading. A swarm-wide Redis channel carries the group chat and coordinating signals.
Why decomposition is the real craft
Splitting a task five ways isn't automatically five times faster. If the pieces overlap, workers step on each other. If the pieces are too fine-grained, the orchestrator burns more time coordinating than the workers save. If the pieces aren't independently executable, the whole swarm stalls waiting on one blocker.
The orchestrator's job is to find the seams. A great decomposition has independent pieces, clear output contracts between them, and a team size matched to the task — two workers for a research brief, four for a competitive analysis, six for a full launch plan. Swarm Mode is built around this judgment.
“Spawn fast, stay in the loop. Workers are autonomous — give them room to execute, and synthesize when they're done.
Where it earns its keep
Swarm Mode pays off the most on tasks that are naturally parallel and judgment-heavy. Things like:
- Tender evaluations — one worker per bidder, reading every attachment, scoring on your commercial preferences, all in parallel.
- Competitive landscape reports — one worker per competitor, each producing a full strategic brief, merged into a single comparison deck.
- Multi-vendor RFQ assembly — one worker drafts the spec, others pull vendor history, pricing, references, and compliance.
- Due diligence packs — workers split across financials, legal, technical, market, and team, each producing a section of the memo.
These are the tasks where an individual agent gets slow, not because the model is slow, but because the work is inherently wide. Swarm Mode makes the width visible — and parallelizable.
What's shipping today
- One-prompt swarm creation — describe the job, the orchestrator handles team shape and coordination.
- Live swarm dashboard — watch every worker's status, message stream, and artifact output in real time.
- Group chat and direct messaging — talk to the whole team or redirect an individual worker.
- Shared filesystem — workers coordinate on files without re-uploading.
- Up to 12 parallel workers per swarm on Scale, unlimited on Enterprise.
What's next
We're working on persistent swarm templates — reusable team shapes you save after a run and re-spawn on a cron or webhook. Paired with Spawn Workflows, a templated swarm becomes a single graph node that expands into five parallel agents whenever it fires. That's the shape we think most serious back-office automation ends up looking like.
Swarm Mode is live today on every Scale and Enterprise plan. Try it at /new — pick 'Swarm Mode' in the composer and watch your team spin up.