Swarm coding is one person running many AI agents at the same time. That person picks the goal, the checks, and when to stop. The agents do the writing, often in parallel. It is a real way to work. It is not a community.
The short line this site uses:
- Vibe coding is one human and one AI. Compared in Agentic swarm coding vs vibe coding.
- Swarm coding is one human and many AIs. This page.
- Hive coding is many people and many AIs. Defined in What is hive coding?.
Who decides in the neighboring product modes, including Cursor multiplayer and Antigravity's other commands, is a separate guide: Swarm vs multiplayer vs hive coding. This page stays on swarm coding. Words for the modes are in the glossary.
Two things people mean
A practice: one person, many agents
People were running several coding agents before the name settled. One chat plans. Other sessions edit different parts of the repo, or race the same problem. The person watches, throws out the bad attempts, and keeps a result.
On November 10, 2026, Matt Huang published "Swarm Riders and Swarm-Native Companies". He used "swarm riders" for people who can consistently point a large swarm of agents at useful work, and he called that a new kind of management skill. The essay says it was first posted on X the same day. The glossary entry is swarm rider.
A swarm rider is still one person. The agents do not take requests from people outside that person's own list. Huang's essay is about companies organized around that skill. It is not a public suggestion queue.
A product command: /swarm
/swarm is a command in Google Antigravity. The docs, checked on November 16, 2026, say your current conversation becomes a coordinator. It names workers, gives each one a public brief, and usually adds a verifier for every two to four workers. The verifier reads the diffs, runs the tests, and reports problems back to the worker that wrote the code. Everyone talks on one shared board, not in private. Idle workers wait without using tokens until someone mentions them.
You type the command and describe the task, including how to check it. You can steer the coordinator while it runs. When it finishes, it reports to you. Workers use the same model as your conversation unless you ask for a different one. The docs' example of a different model is the verifier.
Swarm is in preview. It is on Google AI Pro and Google AI Ultra, in the app and in the CLI. The docs also say a swarm uses quota much faster than one conversation, because every worker is its own agent. Docs: Swarm. The glossary line is /swarm.
That is swarm coding with a button. One person decided the task.
What it is good at
Swarm coding fits work one person already understands, when the work splits.
- Several files or packages can change at once, without two agents editing the same lines.
- A question can be investigated on a few tracks, then combined.
- A failing test suite can be split by package, with a verifier running the full suite at the end.
- One person's afternoon can cover more ground than one chat would.
The verifier is a real help. It is the first reader, not the last one. Why an auditor from the same setup shares blind spots is in Who verifies the verifier?.
Where it stops
One person decides what is worth building. A swarm is only as good as the task it was given. It will not hear from the barber who needs the booking form to remember a phone number, unless that person is the one at the keyboard or someone the rider already listens to.
The checks can be gamed by the same run. If the agents can edit the tests or the benchmark, a green result can mean the check got easier. Fix the check first. Then look yourself.
It spends the quota. Antigravity says so on the swarm docs: every worker is its own agent. A wide swarm on a vague prompt is an expensive way to get a pile of diffs.
There is no public credit for people who were not in the room. The history shows the person who ran the agents. It does not show a community that asked.
How hive coding differs
Hive coding uses agents too. It changes the seat that says what to build.
Anyone can post a suggestion in public. People vote. A maintainer accepts or declines, with a reason. An agent the maintainer connects can open a pull request. A person reviews and merges. The person who asked gets credit. That loop is How the hive coding loop works.
A swarm rider can be the maintainer. The swarm can be how the accepted suggestion gets written. The community is what makes it hive coding. Without the open queue, it is swarm coding with a nicer name.
The Hive Code is one home for the open queue. It is live and just opened. On November 16, 2026 the public idea board still said "No ideas yet. Post the first one." There is no crowd and no ship history here to point at yet.
Frequently asked questions
What is swarm coding?
One person running many AI agents at the same time. That person decides the goal, how to check the work, and when to stop. The agents write. Swarm coding is one human and many AIs. Hive coding is many people and many AIs.
What is a swarm rider?
A person who can keep a large group of coding agents pointed at useful work. Matt Huang used the name in a public essay on November 10, 2026. It is swarm coding described as a skill. One person decides.
What is Antigravity /swarm?
A command in Google Antigravity. The docs, checked on November 16, 2026, say it turns your conversation into a coordinator that splits one task across named workers and a verifier on a shared board. You start it. It is in preview, on Google AI Pro and Ultra.
What is swarm coding good at?
Work that splits across files, packages, or questions, when one person already knows the goal. Several agents can move at once, and a verifier can run the tests before that person looks.
How is swarm coding different from hive coding?
Swarm coding answers how one person coordinates many agents. Hive coding answers who decides what is worth building: an open community, in public, with a person merging. The agents can be the same. The seat that decides is not.
One person is a real start
If the work is yours and it splits cleanly, a swarm is a fair tool. If people who are not at your keyboard should be able to say what gets built, the next step is a public queue and a person on the merge.
Sign up free and post the first idea
Keep reading: