In the multi-agent coding modes that shipped this season, one person or one invited team decides what gets built. Hive coding is the open one: a community decides what is worth building, agents build it, and a person reviews and merges.
The short line this site uses:
- Vibe coding is one human and one AI.
- Swarm coding is one human and many AIs.
- Hive coding is many people and many AIs.
"Team" is doing two jobs right now, so this page keeps them apart. An agent team is several agents with roles. A human team is several people. Google's /teamwork is an agent team that one person starts. Cursor multiplayer is a human team, and each person has agents. Hive coding is neither a solo rider nor a hired team. The people are whoever shows up, in public.
Every product claim below was checked against the linked public page on November 16, 2026. Plans and buttons change. If a doc and this page ever disagree, trust the doc. Nothing here is a paid placement. Words for the modes are also in the glossary.
Who decides
| Mode | Who decides what gets built | Who writes | Who checks before it counts |
|---|---|---|---|
| Vibe coding | One person | One AI | Often nobody |
| Antigravity subagents | One person | A background agent, on one side task | The parent agent, then the person |
| /boost | One person | A pipeline of workers, on one hard problem | Workers try to break the attempt, then the person |
| /swarm | One person | Named workers | A verifier, then the person |
| Swarm rider | One person | Many agents, as a practice | That person |
| /teamwork | One person, after a scoping interview | An agent team with roles | Review gates, then the person |
| Cursor multiplayer | The invited team | Each member's own agents | The people on that team |
| Hive coding | An open community, by public suggestions and votes | A coding agent the maintainer connects | A person merges |
The tools in the middle of that table are good at a real problem: several agents stepping on each other. They answer how the agents coordinate. They do not hand the "what should we build?" question to people outside the room.
One person
Subagents
In Google Antigravity, a subagent is for a focused side task, such as reading the codebase, checking the browser, or running tests. The docs say it runs in the background and reports to the agent that started it. The parent agent checks the reply. You can let the main agent delegate, or mention a subagent yourself. Subagents are on all plans. Docs: Custom subagents.
That is one person, one conversation, and a helper. It is the smallest step past a single agent.
/boost
/boost is for a problem that is hard rather than wide: a tough bug, a race, an algorithm, a refactor where the pieces cannot be split apart. The docs say it starts immediately and runs a pipeline. Workers implement, challenge, fix, and check again. It is on paid plans. Docs: Boost.
You still bring the problem. The command spends more agents on solving it.
/swarm
/swarm is for work that splits cleanly across files, packages, or questions. The docs say your conversation becomes a coordinator. It names workers, gives each one a public brief, and usually adds a verifier that reads the diffs and runs the tests. The workers talk on one shared message board, not in private. Idle workers wait without using tokens until someone mentions them. Swarm is in preview, and it is on Google AI Pro and Ultra. The docs also warn that a swarm uses quota faster than one conversation, because every worker is its own agent. Docs: Swarm.
You type the command and describe the task. You can steer the coordinator while it runs. When it finishes, it reports to you. One person decided the task.
Swarm riders
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.
That is a name for the one-person pattern, not a product. A swarm rider decides the goal, the checks, and when to stop. The agents do not take requests from an open community. The glossary entry is swarm rider.
One team
/teamwork is an agent team, not a human team
/teamwork is Google's command for a long project: a migration, a simulation, a proof, something that can run for hours or days. The docs say it starts with a scoping interview. You approve a written brief. Then a coordinator called Sentinel and a project orchestrator run specialized agents through milestones. Review gates include a Critic, a Challenger, an Auditor, and a final Success Auditor. The work sits in its own folder, and workers are not supposed to edit the same file at once. It is on paid plans. Docs: Teamwork.
Read the word "team" in that doc as agents. One person still approves what the run is for. Strangers cannot file a suggestion into someone else's /teamwork session.
Cursor multiplayer Projects
On November 10, 2026, Cursor announced that Projects is multiplayer (the announcement, Aryaman Khandelwal). Projects already let one person direct a coordinator that plans work and hands it to worker agents. Multiplayer invites teammates into that Project. Each person gets their own chat, their own coordinator, and their own workers. Their chat bills as them. Everyone shares the task board and a shared context store. Coordinators can message each other. Cursor's post says the finished work still comes back for a person to check.
The docs, checked the same day as this page, add limits the blog post does not spell out. Sharing a Project needs a team account and Cursor 3.26.13 or later. Joining takes an invite. Someone on a personal account, or on another team, cannot join. A side chat that answers another teammate's agent is not a place you type, and the docs say it does not share secrets. Docs: Projects.
So multiplayer is a real change. Agents that used to serve one person can now serve a team, with a shared board. The team is closed. The people who were invited decide. An open community does not.
On November 12, 2026, The New Stack put those two launches side by side: Antigravity /swarm for one developer, and Cursor multiplayer so teammates' agents can talk (Paul Sawers, The New Stack). That article is a news report of the week. The product pages above are the source for how each one works.
Hive coding
Hive coding keeps the agents and changes who is allowed to say what is worth building.
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. The full loop is in How the hive coding loop works. The definition is in What is hive coding?.
That is many people and many AIs. It is not a livestream chat, and it is not a company Slack. The queue is open. The merge is not. Community text reaches an agent only as a labeled, quoted request a human already approved. The risks, and the rules that answer them, are in Is hive coding safe?.
The Hive Code is one home for that loop. 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 to point at yet. Other sites that describe people and agents working together are compared, from their own pages only, in Best hive coding platforms, compared (2026).
Liquid software
On October 31, 2026, Patrick Collison described a personal pattern he called liquid software: a text box on an app he vibe coded, hooked to an agent that implements whatever someone submits (his note). He wrote that the pattern works well for him, and he asked the questions that matter if it leaves one person's apps. Does it become unmaintainable? How does collaborative software work? Can it be secure and reliable?
Those questions are the gap between a personal agent box and hive coding. A community can decide in public. A person still reviews before a change ships. The short definition is in the glossary under liquid software.
How to choose
Pick from who should decide, not from which word is loudest this week.
- The stakes are low and the audience is you: vibe coding is enough.
- You have one hard bug:
/boostis built for that. - You have one wide task that splits across files:
/swarm, or your own swarm of agents if you are the rider. - The work will run for days and you want milestone checks:
/teamwork. - A known team already shares the product: Cursor multiplayer Projects.
- People who are not on that team should be able to say what gets built, and a person should still merge: hive coding.
Frequently asked questions
Who decides what gets built in hive coding?
An open community. People suggest and vote in public, a coding agent writes the approved change, and a person reviews and merges it. In the other modes on this page, one person or one invited team decides.
What is the difference between vibe coding, swarm coding, and hive coding?
Vibe coding is one human and one AI. Swarm coding is one human and many AIs. Hive coding is many people and many AIs: a community decides what is worth building, agents build it, and a person reviews and merges.
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 with a name. One person decides.
What is the difference between /swarm and /teamwork?
Both are commands in Google Antigravity, and both are started by one person. /swarm splits one task across named workers and a verifier on a shared board. /teamwork is for a longer project: a scoping interview, then an agent team with milestone review gates. The team in /teamwork is agents, not a group of people.
What is Cursor multiplayer?
Cursor's November 10, 2026 upgrade to Projects. Teammates you invite share a task board and shared context, and each person keeps their own chat, coordinator, and worker agents. The docs checked on November 16, 2026 say joining takes a team account and an invite.
Is /teamwork the same as a human team?
No. Google's docs describe /teamwork as an agent team: specialized agents with roles, started after one person approves a brief. A human team is Cursor multiplayer, where several people each bring their own agents. An open community is hive coding.
The open seat is still empty here
The other modes already have products. The community seat is what The Hive Code is for, and it just opened.
Sign up free and post the first idea
Keep reading: