Google Antigravity's /teamwork needs a paid plan and can flood your folder with files. Claude Code subagents are cheaper but need manual orchestration. Here's which wins for long projects.
Vamsi Tallapudi
Manager, Architect Technology at Cognizant
If you're picking between Google Antigravity's /teamwork and Claude Code's subagents for a project that'll take days instead of minutes, here's the short version: /teamwork is the heavier tool built for genuinely long, multi-milestone work and it needs a paid plan ($19.99/month Pro at minimum), while Claude Code subagents are lighter, cheaper to run since they just count against your existing plan, and easier to steer, but you have to design the orchestration yourself. Neither one is "done" software yet. Both have rough edges that show up fast once you push real work through them.
I've spent the last couple weeks running both on actual refactors, not toy demos, and the gap between what Google's docs promise and what Reddit/Discourse users are reporting is wide enough that it's worth walking through before you commit a real project to either one.
/teamwork spins up a full agent team instead of one model. Per Google's own docs, it's built for "large software projects, complex simulations, and deep research tasks" that run for hours or days, not minutes.
The team has eight named roles split into three tiers:
Every project gets its own folder at ~/teamwork_projects/{PROJECT_NAME}, and no two Workers touch the same file at once. You pick one of three "integrity modes" (development, demo, benchmark) up front, which controls how strict the Auditor is about fabricated test results or copy-pasted solutions.
The workflow is two phases: first a scoping interview where Sentinel asks about requirements and acceptance criteria and hands you a prompt artifact to approve, then an autonomous execution phase where the team just runs. You don't babysit step by step. You approve the plan, then come back later.
Subagents in Claude Code are closer to specialized contractors than a standing team. You define them as markdown files with YAML frontmatter, either project-scoped in .claude/agents/ or global in ~/.claude/agents/, and the main Claude session (the orchestrator) decides when to call them through the Task tool.
The core difference from /teamwork: there's no scoping interview, no fixed roles, and no dedicated project folder. You just prompt, and Claude either builds a plan itself or you tell it explicitly to fan work out across subagents. Each subagent gets its own context window, does its work, and returns a summary, which is the whole point, since it keeps your main session's context from filling up with exploration noise it doesn't need to remember.
For genuinely long multi-day work, the piece that matters most isn't subagents on their own, it's that Claude Code shipped native git worktree support on February 21, 2026. Run claude --worktree task-name and you get an isolated branch and directory sharing the same .git store, so two or more full Claude sessions (not just subagents) can work the same repo without stepping on each other's files.
Google Antigravity /teamwork | Claude Code Subagents | |
|---|---|---|
| Minimum cost | Paid plan required ($19.99/mo Pro, up to $199.99/mo Ultra) | Included in your existing Claude Code plan |
| Setup | Scoping interview + approved prompt artifact before anything runs | Just prompt; no upfront interview |
| Roles | 8 fixed roles across 3 tiers | You define roles in markdown, as many or few as you want |
| File isolation | Exclusive file ownership enforced by Orchestrator | Manual, unless paired with git worktrees |
| Typical horizon | Hours to days | Minutes to hours per subagent call, days if chained manually |
| Verbosity | Can generate dozens of markdown status files | Returns a summary, discards the rest |
| Best for | Repo-scale migrations, research tasks, anything with a hard verification bar | Parallel exploration, code review, context isolation on one big task |
The Hacker News thread on Google's teamwork results is a useful reality check against Google's own framing. Commenters generally agree multi-agent, role-based setups beat single-agent runs on hard problems, that part isn't controversial. But the complaints are specific: the harness asks for approvals constantly even in "auto mode," and more than one person called it rough around edges that would be trivial for Google to fix.
The discussion on Google's AI developer forum is worse. One user on the Ultra plan ran a real computational task and got back roughly 50 markdown files cluttering their working directory, plus a token bill they described as "way too much uncontrolled exhaust." Another complaint: the team spent most of its run spinning up overly aggressive test suites instead of making progress on the actual task, something that didn't happen when the same prompt went through Antigravity's normal single-agent pipeline.
That tracks with what Google admits in its own docs: /teamwork's horizon is measured in hours to days, specifically because it resets context between milestones to avoid degradation. That's a reasonable design choice for a six-day research problem. It's a bad fit if you just wanted a quick multi-file refactor, which is exactly why Google also ships /boost (hard but narrow problems) and /swarm (broad problems that split cleanly across files, with a shared message board instead of a hierarchy) as lighter alternatives.
Subagents solve context isolation well, but they don't solve coordination for you. If your subtasks depend on each other, firing them in parallel breaks things, and there's no built-in scoping interview or milestone structure forcing you to think that through first. You're the Project Orchestrator here, whether you wanted the job or not.
The other real cost: subagent requests count toward your same Claude Code usage limits, not a separate pool. Aggressive delegation, spinning up five subagents for a task one could've handled, burns through your weekly limit faster and adds latency waiting on the slowest one. I've had sessions where I fanned out subagents I didn't need and just paid for it in token budget.
For a genuinely long project, repo-scale migration, a research task with a hard correctness bar, something you're fine kicking off and checking back on in a day, /teamwork's structure (forced milestones, adversarial review gates, exclusive file ownership) is doing real work that you'd otherwise have to build yourself. Just budget for the paid plan and expect to tell it explicitly to keep output contained, because left alone it will generate more files than you want.
For everything else, which is most actual engineering work, Claude Code subagents plus native git worktrees get you 80% of the benefit without the $19.99-to-$199.99 monthly floor or the mandatory scoping interview. You design less structure, but you also don't inherit structure you didn't ask for.
Honestly, if your team already pays for Claude Code, start there and only reach for /teamwork when a task is big enough that you'd otherwise be managing the milestones by hand anyway. We wrote more about Antigravity's broader agent setup in our piece on Gemini 4 Hydrogen and the Antigravity orchestrator, and if you're trying to figure out where Claude Code sits against Codex and Cursor more broadly, this comparison is a good next read. For the cost side of running subagents hard, see our breakdown of Claude Code usage limits and pricing.
It spins up a team of eight specialized agents (Sentinel, Project Orchestrator, Explorers, Workers, Critic, Challenger, Auditor, Success Auditor) that plan, build, and verify large software or research tasks over hours to days, with exclusive file ownership so Workers never edit the same file at once.
Yes. /teamwork only runs on paid Google Antigravity plans, starting at Pro for $19.99/month and going up to Ultra tiers at $99.99 and $199.99/month. It's not available on the free tier.
Subagents are lighter weight: you define them yourself in markdown files, there's no forced scoping interview or fixed role structure, and they run on your existing Claude Code plan instead of requiring a separate paid tier. /teamwork trades that flexibility for built-in milestones and adversarial review gates.
Subagent requests count toward the same usage limits as your main Claude Code session, so there's no separate fee, but heavy delegation still eats your weekly quota faster since each subagent call adds its own tokens and latency.
/teamwork is built specifically for that horizon, with forced milestones and file-ownership rules that prevent agents from colliding. Claude Code can match it if you pair subagents with native git worktrees (available since February 21, 2026), but you have to set that structure up yourself.
/boost is for problems that are hard but narrow, like a tricky race condition, and runs a single deep reasoning loop. /swarm is for broad tasks that split cleanly across files, with workers coordinating on a shared message board instead of a hierarchy. /teamwork is the heaviest of the three, meant for projects that genuinely run hours to days with planning and review gates built in.
New AI tools, automation workflows, and course drops — straight to your inbox. Join 2,400+ builders.
Pinky is a free open-source MCP gateway from DIYgod that lets Claude Code or Codex control your real iPhone or Android phone with a human approval gate for risky actions.
Qwen4-Max Preview packs a 1M-token context and agent-first coding claims, but Alibaba shipped it with no benchmarks and a hard November 25 shutoff. Here's what's real.
Expo's new Exact framework runs one Contract app natively on web, macOS, iOS and Linux in a ~20KB runtime, built for AI coding agents to write, run and test it.
Comments