Agent Manager in Antigravity: coordinating multiple agents in parallel

November 3, 2026

Updated:

Guide
Agent Manager in Antigravity: coordinating multiple agents in parallel

What is Agent Manager in Antigravity?

Agent Manager is the coordination layer of Google Antigravity, where one person opens multiple conversations with agents at the same time, tracks each agent, reviews permission requests and steps in to correct course when an agent goes off track. According to the official documentation (read 03/11/2026), the Antigravity application "evolves from Agent Manager" into a standalone desktop application, while in Antigravity CLI the /agents command opens a panel named Agent Manager. This post does not promise that running in parallel will be X times faster: OPCAIOS has not yet measured productivity. The post goes through how to delegate work, track progress, intervene, child agents, /teamwork, and when parallelization helps or causes trouble. The general introduction is in the post What is Antigravity?.

Diagram of Agent Manager in Antigravity: multiple conversations in one project, Local or New Worktree mode, main agent calling child agents, states running, idle, killed, permission requests bubbled up to the main interface, and /teamwork with coordinator, coder, and reviewer roles
Agent Manager: multiple conversations, child agents and agent teams all in one place to track. Diagram: OPCAIOS.

Where is Agent Manager in Antigravity?

Agent Manager in Antigravity currently appears in three places with different names. According to the official changelog, the first release on 18/11/2025 was an IDE with a new "Agent Manager view". The Antigravity agent page still mentions Agent Manager and opening multiple conversations in parallel. The Antigravity application is described in the overview documentation (read 03/11/2026) as a place to launch, coordinate and review agent teams, with a sidebar divided by project, conversation history and Automations section. Antigravity CLI version 1.4.12 has the /agents command that opens the Agent Manager panel to select custom agents and monitor background child agents.

How do I delegate multiple tasks at once in Agent Manager?

Delegating multiple tasks at once in Agent Manager has two ways: open multiple conversations yourself, or let the main agent split the work to child agents. The first way works when the tasks are independent and you want to manage each one yourself. According to the Projects page (read 03/11/2026), each new conversation has a worktree selector with two modes:

  1. Local mode: the agent works directly on the folder or Git checkout you have open, good for quick fixes.
  2. New worktree mode: creates a new Git worktree for the conversation, keeps the main folder untouched and prevents parallel agents from colliding.

One project can consist of multiple folders, for example one frontend repo and one backend repo, and each has its own permission settings. The second way is to give a big task and let the main agent call child agents. Example scenario: a web developer opens three conversations, one writes tests in a new worktree, one reads old code to write documentation, one fixes UI bugs in Local mode.

How do I track and intervene when multiple agents run at the same time?

Tracking in Agent Manager is seeing the status, current step and pending approval requests from each agent, then stepping in at the right place. According to the Custom subagents page (read 03/11/2026), on the application you open Subagents in the Overview panel to see each child agent's conversation. To change instructions without going through the main agent, you type directly into the child agent's message box. On CLI, the /agents panel shows the code, role, status and current step of each agent. Four main intervention actions:

  1. Stop a child agent: Stop Subagent button on the app, K key in CLI /agents.
  2. Approve permissions right in the panel: A or D key; Alt+J jumps to the agent waiting for approval; Ctrl+K quick approve.
  3. Send a private message to a child agent on CLI: start with @ and the agent name.
  4. Track background tasks not tied to an agent using /tasks.

When a child agent needs permission, the request bubbles up to the main interface, so you don't have to open each window to find it.

How do subagents in Antigravity work?

A subagent in Antigravity is a parallel work session called by the main agent through the invoke_subagent tool, with its own role and starting with an empty context. According to the Custom subagents page (read 03/11/2026), a subagent has three workspace options: share a directory with the parent agent (inherit), create its own worktree (branch), or share a repository (share). Antigravity comes with three built-in subagents: research to survey code, browser to test the browser via /browser, and self which is a copy of the calling agent. You can define your own subagents using a Markdown file with YAML frontmatter in .agents/agents/ of your project, with keys like name, description, tools, model, and commandExecutionPolicy. A subagent has three states: Running, Idle, and Killed. An Idle agent will restart itself when it receives a message and still remembers previous work. The maximum nesting level is 10 layers. A subagent inherits the command permissions, directory scope, and sandbox of its parent agent. Known issue: misspelling a tool name in tools can cause the subagent to hang. For the general concept, see What is a subagent?.

What's the difference between /teamwork and /boost versus opening multiple agents yourself?

/teamwork is a team agent mode for projects lasting hours to days, while /boost is a multi-agent reasoning chain for a difficult problem within a single session. According to documentation read on 03/11/2026, both are only available on paid plans. The key difference is that you don't divide the work yourself. /teamwork has two phases:

  1. Interview phase: the main agent asks about scope, requirements, verification methods, acceptance criteria, then creates a prompt artifact for you to review.
  2. Auto-run phase: Sentinel coordinates, Project Orchestrator breaks down milestones, Explorers read-only, Workers write code; Critic, Challenger, Auditor, and Success Auditor check independently.

A point worth learning is the file ownership rule: Orchestrator assigns each file to each Worker, and two Workers never edit the same file at the same time. According to the Boost page, /boost runs in a temporary worktree, while /teamwork uses a dedicated worktree for each milestone.

When is running multiple agents in parallel beneficial?

Running multiple agents in parallel is beneficial when tasks are independent of files, have clear completion criteria, and you have enough time to review each result. Four reasonable cases, drawn from how Antigravity documentation describes subagents and /teamwork rules:

  1. Read-only survey: the research subagent reads code, traces function calls, doesn't write files so no conflicts.
  2. Heavy background work: run tests, builds, code scans while you continue other work.
  3. Isolated modules: each agent edits a separate directory in its own worktree.
  4. Cross-checking: one agent writes, another reads only to catch errors, like the Critic role.

Hypothetical example: a shop owner has one agent write a returns policy page, one subagent checks for broken links across the site. With long work, the /goal command keeps the agent from stopping early and still waits for the subagent to finish.

When does parallel running cause trouble: file conflicts and quota drain?

Running in parallel causes trouble when multiple agents touch the same file or resource, or when total work exceeds quota and exceeds one person's review capacity. Five common risks:

  1. File conflicts: two agents in Local mode edit the same file, the later version overwrites the earlier one. Mitigation: New Worktree mode and the one-file-one-owner rule.
  2. Shared resources: run the same test server, migrate the same database, commit together. Worktree cannot isolate a database.
  3. Quota drain: according to the Plans page (read 03/11/2026), usage limits correlate with work volume, complex work costs more. Free plan refreshes weekly; Google AI Pro and Ultra refresh yearly with hourly limits until hitting weekly cap.
  4. Review bottleneck: five agents finish at once means five sets of changes to read.
  5. Permission creep: a subagent inherits the parent agent's permissions, so Turbo mode on the main agent expands to the whole tree.

For threading multiple sessions on one repository, see multiple AI sessions editing one project; for permissions, see safety modes in Antigravity.

OPCAIOS Experience

In the OPCAIOS system, Antigravity acts as a coordination agent on a Mac: managing infrastructure, dispatching and tracking work. This machine has Antigravity app version 2.28.x, Antigravity IDE version 2.5.x and CLI agy version 1.4.12, running alongside Claude Code and Codex (see using Codex and Claude Code together). Antigravity has handled some website administration features, such as tracking Dream 100 lists on Facebook and X, and editing menus. Four lessons learned:

  1. Website code is only edited and deployed from a single machine, after encountering over 1,500 different files due to line break differences when deploying from a Windows machine.
  2. A shared AGENTS.md file specifies the order of reading documentation at session start and logging status at session end via a shared state file.
  3. The machine has over 400 saved Antigravity conversations stored locally (counted by directory, not by content), requiring clear naming and regular cleanup.
  4. Hooks currently only report session status to one terminal management tool; there are no command-blocking hooks yet.

We have not used Antigravity SDK, have not used Remote Control, do not have enough experience to evaluate /teamwork or /boost, and have no productivity metrics.

Frequently Asked Questions

Is Agent Manager in Antigravity free?

Opening multiple conversations and using subagents is available on all plans: the Plans page (read 03/11/2026) states that all plans can use product features. Only /teamwork and /boost are available on paid plans. The free plan has a weekly refresh rate limit, and because the limit correlates with work volume, running many agents at once will deplete it faster.

Should I choose Local mode or New Worktree mode?

Choose Local mode when only one agent edits files, when agents only read, or when small tasks need immediate visibility in the open directory. Choose New Worktree mode when two or more agents write files in one Git repository, or when large tasks need a clean main directory. Worktree only separates source code; databases, test ports, and files outside the repository remain shared, so work still needs careful division.

Do subagents remember the main agent's conversation history?

No. According to the Custom subagents documentation, a subagent starts with empty context and only receives the role and work instructions given by the main agent. This keeps context lean, but work instructions must be complete: goals, file scope, completion criteria. When a subagent finishes and moves to Idle, sending more messages will wake it, and it will still remember the previous work.

How do I stop a subagent that is running too long?

On the Antigravity app, open the subagent in the Subagents section and click Stop Subagent. On CLI, type /agents, select the running line and press K; this action stops all subagents under it. A stopped agent cannot be woken again, but the conversation record can still be read.

Should a beginner run many agents in parallel right away?

No. Start with one agent, one task, approval-required permission mode, and learn to read plans and changes before accepting (see Artifacts in Antigravity). Once comfortable, add a read-only subagent for exploration, then move to multiple agents writing files in separate worktrees. To learn a controlled process for using programming agents, see OPCAIOS AI application training.

Source: Google Antigravity documentation at antigravity.google/docs: overview, features, agent, subagents, teamwork, boost, projects, plans, commands/goal, cli/commands/agents, at-mentions, agent-settings, changelog (read 03/11/2026); CLI flags checked using agy --help version 1.4.12. Not verified: when Agent Manager was removed from IDE, actual effectiveness of /teamwork and /boost, specific rate limit consumption when running in parallel. Experience is from OPCAIOS.

Translated from Vietnamese with AI. Read the original.

Want to apply this in your business?

Book a free 30-minute call (Vietnam time, in English or Vietnamese).

Book a call