Mirasim
Get Mirasim
← Blog

How-to4 min read

Claude Code agent teams, and teams that mix agents and people

Claude Code's agent teams let a lead session split work across Claude teammates. How to turn them on, what they cost, and how Mirasim mixes Codex and people in.

Two Mirasim sessions side by side, one handed over from Claude Code to Kimi CodeTwo Mirasim sessions side by side, one handed over from Claude Code to Kimi Code

The short answer

Agent teams are an experimental Claude Code feature from Anthropic, off by default. One session acts as the team lead: it splits the work, assigns tasks and pulls the results together. The teammates are other Claude Code sessions, each with its own context, sharing a task list and messaging each other directly.

  1. Turn it on

    Set CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 in settings.json or the environment

  2. Ask for a team

    In an interactive session, the lead spawns teammates and hands out tasks

  3. Split the files

    Teammates share one checkout, so give each its own part

Anthropic notes that agent teams use significantly more tokens than a single session, and work best when teammates can operate independently.

Claude Code's five ways to run agents in parallel

What it isUse it when
SubagentsWorkers inside one session that report a summary backA side task would flood your main conversation
Agent viewOne screen to dispatch and watch background sessions (claude agents); research previewYou have several independent tasks to hand off
Agent teamsA lead plus teammates with a shared task list and messaging; experimental, off by defaultClaude should split a project and keep the workers in sync
Dynamic workflowsA script that runs many subagents and cross-checks their resultsThe job is too big for a handful of subagents
ProjectsOne ongoing conversation at claude.ai/code or in the desktop app that starts parallel threadsYou want Claude to coordinate a longer body of work

In every one of these, the workers are Claude sessions. Anthropic's guide points to MCP for bringing another tool in.

Teams that mix agents and people, in Mirasim

A Claude Code sessionName another session with @, or let the agent start one
  • A Codex sessionStarted with the agent and model the task needs
  • A Kimi Code sessionRead where it has got to, or send it a message
  • A teammate's sessionIn a team space, people steer sessions too
Session messaging works across agents: one session's agent can read another session, message it, or start a new session with a chosen agent, model and opening task.
  • Each parallel session can start in its own git worktree, so sessions don't overwrite each other's files.
  • In a team space, several people send into the same session, each message signed with its sender.
  • Every session, whichever agent runs it, is in one sidebar.

Chapter 8 · Session messaging →

Questions

Are agent teams on by default?
No. They're experimental; set CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS to 1 to turn them on. Teammates are only spawned in interactive sessions, not with -p.
What's the difference between subagents and agent teams?
Subagents work for one session and report a summary back. Agent teams are separate sessions that share a task list and message each other, which costs more tokens.
Can an agent team include Codex?
Not in Claude Code: its teammates are Claude sessions. In Mirasim, a Claude Code session can start or message a Codex session through session messaging.
Do teammates get their own worktree?
Anthropic's guide says agent teams don't isolate teammates in worktrees, so split the work so each teammate owns different files.

Article · Running two coding agents in one repository → Running parallel agents in separate git worktrees.

Sources

Agent teams and Claude Code's other parallel options were checked against Anthropic's documentation on November 8, 2026: code.claude.com/docs/en/agent-teams and code.claude.com/docs/en/agents. The feature is experimental and changes quickly, so check those pages for the latest.

Keep reading