Home › Rules › Antigravity /boost vs /teamwork

Google Antigravity /boost vs /teamwork: when to use which

From Google's own Antigravity docs (Boost deep reasoning, Teamwork agent teams, and Slash commands), checked Nov 7, 2026. Sources below.

• Guide
• Antigravity
• /boost
• /teamwork
• Beginners

Short answer: /boost and /teamwork are Antigravity slash commands that start multi-agent work for you. Use /boost when a tough bug, race condition, or tricky refactor needs deeper reasoning inside a normal coding session (seconds to hours). Use /teamwork when the job is large enough to need a multi-day agent team, a scoping interview first, and milestone checks (hours to days). Both are on paid plans in the Antigravity App and the Antigravity CLI. Everyday coding still uses the default agent on all plans.

Default vs /boost vs /teamwork

Google's Boost doc compares the three primary execution modes. Dimensions below are Google's:

Google's slash-commands catalog says the same plan and horizon notes: /boost and /teamwork are both "Paid plans," with horizons of seconds-to-hours and hours-to-days.

Plan availability

• /boost: "available on paid plans across the Google Antigravity App and the Antigravity CLI" (Boost doc).
• /teamwork: same wording on the Teamwork doc — paid plans, App and CLI.
• Default agent: all plans (from the comparison table).

Google does not list free-tier access for either command on those pages. If you are on a free plan, use the default agent, or see Antigravity limits and AI credits for how quotas and paid tiers work.

How /boost works (3 phases)

Google describes /boost as an on-demand multi-agent reasoning pipeline for challenging software engineering tasks. When a single-turn session is not enough for a complex bug, race condition, or intricate refactor, Boost breaks the problem down, hands focused workstreams to specialized subagents, and verifies solutions across iterative rounds.

1. Phase 1 — Goal and strategy formulation: the primary orchestrator reads your prompt, inspects workspace context, forms a strategy, splits the work into discrete verifiable subtasks, and decides which specialized workstreams are needed.
2. Phase 2 — Parallel execution and verification: the orchestrator dispatches subtasks to specialized subagents in clean, isolated scopes. Implementation workstreams build candidate code, apply refactors, and write unit tests. Investigation workstreams do root-cause debugging and call-graph tracing without modifying files. Subagents run builds and tests locally before reporting back.
3. Phase 3 — Synthesis and delivery: findings are aggregated and regression-checked. If an assertion fails, diagnostics go into the next iteration. When tests and requirements pass, you get a concise summary with verified changes.

The slash-commands overview also summarizes Boost as a three-tier hierarchy (orchestrator → DeepCoder / DeepInvestigator → isolated workers) that explores hypotheses, runs tests, and verifies before returning.

When Google says to use /boost

• Concurrency and race conditions (deadlocks, cache sync bugs)
• Algorithmic problem solving (high-performance algorithms, custom data structures, stress tests)
• Non-trivial refactoring across multiple files
• Deep root-cause investigation in unfamiliar or large codebases (including read-only traces)

How to invoke /boost

In the Antigravity App, type /boost plus your task in any conversation turn:

/boost Investigate the race condition in the session cache and implement a thread-safe fix with tests.

In the Antigravity CLI TUI prompt box:

/boost Optimize the matrix transposition algorithm to use SIMD vectorization and benchmark throughput.

Another Google example for a concurrency bug:

/boost Reproduce and fix the intermittent deadlock in the connection pool during high connection turnover.

Boost safety notes (from Google)

• Subagents inherit your workspace file-access rules and command permission policies.
• Protected commands or edits outside trusted scopes still ask you to approve.
• Subagents run in isolated memory so verbose debug logs do not clutter the main chat.

How /teamwork works

Google describes /teamwork as a collaborative multi-agent team for large software projects, complex simulations, and deep research. It pairs specialized agents, runs independent verification at every milestone, and works inside isolated project directories.

Roles

Orchestration and execution:

• Sentinel: coordinator after you approve the prompt — records the request, routes tasks, posts progress, and spawns the Success Auditor before wrapping up.
• Project Orchestrator: breaks the brief into milestones, coordinates parallel tracks, delegates work, and hands off to a fresh successor between milestones to limit context degradation.
• Explorers: read-only research agents (no source edits).
• Workers: implementation agents with terminal and file tools, on non-overlapping tracks.

Adversarial verification gates (before a milestone is approved):

• Critic: independent code review (correctness, completeness, robustness, style).
• Challenger: builds adversarial tests, edge cases, and worst-case inputs.
• Auditor: checks work against the selected integrity mode and real command output (not mocked or skipped tests).
• Success Auditor: final end-to-end pass spawned by Sentinel before the finished project is shown to you.

Phase 1: scoping interview

Teamwork uses a two-phase workflow. Phase 1 is a structured interview with the main Antigravity agent. Google's principle: specify what, not how.

1. Scope and objectives (what to build, purpose, audience)
2. Requirements (only what you care about)
3. Independent verification for each requirement (test suite, benchmark, or rubric)
4. Acceptance criteria
5. Project working directory (default ~/teamwork_projects/{PROJECT_NAME})

You get a reviewable prompt artifact. After you approve it, Phase 2 starts: Sentinel hands off to the Project Orchestrator and the team runs through milestones with structured artifacts (request, project plan, progress).

Integrity modes

During Phase 1, your agent asks which shortcuts are off-limits and maps that to a mode recorded in the prompt artifact:

Execution paths (automatic)

Teamwork picks a path from your prompt. Defaults to General (distributed coding). Saying things like "keep it small" or "keep it focused" can select Iterative coding (one self-contained change; cannot decompose). Review and math/proof prompts select Document Review or Math / Proof paths; "Use a very large team of agents" can select Math / Proof (Large Team).

How to invoke /teamwork

In the Antigravity App, start a new conversation:

/teamwork Migrate our REST API service from Express to Fastify, including full test coverage, TypeScript typing, and benchmark validation.

In the CLI TUI:

/teamwork Design and implement a cycle-accurate memory controller with lockstep simulation tests.

After Phase 1 approval, you can monitor subagents in the UI panel or CLI status bar. Google notes Alt+J jumps to a subagent waiting for approval in the CLI.

Teamwork workspace isolation (from Google)

• Work runs in a dedicated project folder (default ~/teamwork_projects/{PROJECT_NAME}).
• The Project Orchestrator gives each Worker exclusive file ownership so two workers do not edit the same file at once.
• Each subagent has its own scratch directory for helpers and logs.
• State, commands, reviews, and verification logs are recorded in plan and progress artifacts.

Relation to custom subagents

/boost and /teamwork are built-in multi-agent modes Google ships. Custom subagents are different: you write a Markdown agent file (for example under .agents/agents/), and the main agent starts it with invoke_subagent when the description matches. Use custom subagents for repeatable helpers (test runner, read-only researcher). Use /boost or /teamwork when you want Google's packaged multi-agent pipelines for a hard session or a long campaign. Plan quotas still apply to everything agents do — see Antigravity limits and AI credits.

Tips and when not to use them

• Stay on the default agent for everyday features, navigation, and small refactors (Google's "best suited for" row).
• Prefer /boost when you already know the bug or refactor and want deeper multi-agent reasoning without a long interview.
• Prefer /teamwork when the work spans many files or days, needs a written acceptance plan first, or is research / simulation / proof scale.
• Do not expect free-tier access for either command on the docs pages above — both say paid plans.
• Approve Teamwork's prompt artifact before Phase 2. Google designed the interview so you agree on what success looks like before agents write a lot of code.
• Watch quotas. Multi-agent runs use more model calls. Check limits before starting a long Teamwork campaign.

Related

• Google Antigravity subagents (custom agents, agents folder, examples)
• Google Antigravity skills (skills folder, SKILL.md, workflows vs skills)
• Google Antigravity plugins (install, marketplace, plugin vs skill)
• Google Antigravity rules (examples, triggers, where the files go)
• Google Antigravity limits and AI credits (free tier, quotas, AI credits)
• Antigravity vs Claude Code vs Codex (prices, limits, rules files)
• Custom slash commands (how slash commands work across tools)
• Free rules file generator

Sources

• Google Antigravity docs: Boost deep reasoning (/boost)
• Google Antigravity docs: Teamwork agent teams (/teamwork)
• Google Antigravity docs: Slash commands overview
• Google Antigravity blog: Teamwork — When AI Becomes a Research Partner (Aug 27, 2026)

Every claim here was checked Nov 7, 2026 on Google's own Antigravity docs (Boost, Teamwork, Slash commands). Google is still changing these modes. If something here no longer matches, tell us at [email protected].