Opus 4.6 Agent Teams: How Claude Code Teams Work

Anthropic released Claude Opus 4.6 on February 5, 2026, and shipped Opus 4.6 agent teams in Claude Code as a research preview. Opus 4.6 is now a legacy model. Agent teams are still in Claude Code, still marked experimental. A team is several Claude Code sessions working on one job: a lead that hands out tasks and teammates that each own a piece. Below: what the launch claimed, how teams work today per Anthropic's agent teams docs, and what a team remembers once the session ends.

What are Claude Code agent teams, in one sentence?

An agent team is a group of Claude Code sessions in which a lead spawns teammates that coordinate through a shared task list and direct messages.

The feature is still marked experimental and is off by default. You turn it on with one environment variable. Set it in your shell or in the env block of settings.json:

{
  "env": {
    "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"
  }
}

After that you ask for a team in plain language. Anthropic's example asks for three reviewers on one PR, each with a different focus. Teammates need an interactive session. In non-interactive mode with the -p flag, Claude Code does not spawn them.

How do agent teams differ from subagents?

Both split work across context windows. Subagents report only to the main agent, while teammates message each other and you.

Source: Anthropic's agent teams docs.

Anthropic's own advice is to check whether a lighter option works first. The docs say sequential tasks and same-file edits go better in one session or with subagents.

What did the Opus 4.6 launch add?

The Opus 4.6 announcement bundled the model with several platform features. Anthropic described agent teams as agents that work in parallel and coordinate on their own. It pitched them for read-heavy work like codebase reviews.

The model itself kept the Opus 4.5 price of $5 per million input tokens and $25 per million output tokens. It added a 1M token context window in beta. At launch that window was limited to the Claude Developer Platform. Prompts over 200k tokens cost more. Output went up to 128k tokens. Adaptive thinking let the model decide when to reason harder, steered by four effort levels with high as the default. Context compaction, also in beta, summarized older turns as a conversation approached a threshold.

As of October 2026, Anthropic's models overview lists Opus 4.6 among legacy models that are still available. It recommends Claude Opus 5.5 for most work. A team's model is chosen per teammate, so you are not tied to the model that launched the feature.

What do the Opus 4.6 launch claims leave out?

The benchmark numbers are vendor-reported, from Anthropic's own runs. The company reported the top score on Terminal-Bench 2.0 among the models it tested, with the figures in a chart. It reported 81.42% on SWE-bench Verified, but only with a prompt modification and averaged over 25 trials. On GDPval-AA it put Opus 4.6 about 144 Elo points ahead of GPT-5.2. The announcement gives no independent reproduction of these results. It also gives no measurement of how much agent teams improve a task, only the use cases Anthropic expects them to suit.

How do agent teams work?

A team has four parts, per the docs: the lead, the teammates, a shared task list and a mailbox for messages.

The lead is the session you started, and it stays the lead for its lifetime. It breaks the work into tasks. Teammates either receive a task or claim the next unblocked one. Tasks can depend on other tasks, and claiming uses file locks so two teammates cannot grab the same one. Each teammate can message any other by name. You can also open any teammate and talk to it directly. By default all teammates run inside your terminal. With tmux or iTerm2, each one can get its own split pane.

A teammate's model comes from your spawn prompt first. Without one, it uses the named subagent definition, then the lead's model. You can define a role once, such as a security reviewer in .claude/agents/. It then works as both a subagent and a teammate. Hooks add quality gates. A TaskCompleted hook that exits with code 2 keeps a task from being marked done and sends feedback. A TeammateIdle hook can keep a teammate working.

What does a team cost, and where does it break?

Per the docs, token use grows with each active teammate, since each one has its own context window. Anthropic suggests three to five teammates for most workflows and five or six tasks per teammate. It also suggests starting with research and review before parallel implementation.

The docs list these limitations:

• /resume and /rewind do not restore in-process teammates. After a resume, the lead may try to message teammates that no longer exist.
• Task status can lag when a teammate forgets to mark work done, which blocks dependent tasks.
• Each session has exactly one team. Teammates cannot spawn teams of their own, and leadership cannot move to a teammate.
• Two teammates editing the same file overwrite each other, so split work by file.

What does a team remember after the session?

Each teammate loads the same project context as a normal session: CLAUDE.md, MCP servers and skills. It also gets the spawn prompt from the lead. It does not get the lead's conversation history.

When the session ends, Claude Code deletes the team config directory under ~/.claude/teams/. The task list under ~/.claude/tasks/ stays on disk until the cleanup period runs. It holds tasks and their status only. The reasoning behind them lives in transcripts. The findings that teammates argued over end up in the lead's final summary, or in a file you asked them to write. Nothing carries them into the next team, which starts from CLAUDE.md again.

The bottom line

Try agent teams for reviews and debugging with competing hypotheses, where independent viewpoints help and nobody edits the same file. Keep teams small and give each teammate a full spawn prompt. Pre-approve common commands so permission prompts do not pile up on the lead. For sequential work, a single session is cheaper. Decide up front where the team's conclusions should be written, because the team itself will not keep them.

How do you keep what an agent team learned?

Five teammates can debate a bug and agree on the root cause, then their sessions close. The next session sees the task list and none of the reasoning behind it. If part of the work moved to Codex or Cursor, those findings live in yet another place.

memU stores those findings so a later session can retrieve them. A scheduled background task slices new session logs into jobs. The agent decides what is worth keeping and writes it as memory and skill Markdown, which memU commits. memU itself makes no LLM or chat calls. It stores the Markdown the agent prepared and retrieves it by embedding search. A standing instruction in CLAUDE.md tells each new session to retrieve before it answers. Teammates load CLAUDE.md too, so they follow it. All supported hosts share one memory backend, so Codex reads the same wiki as Claude Code.

The hosted version on memu.so is free and syncs across devices. Get an API key there and send Claude Code one message:

Read https://memu.pro/SKILL.md, follow its instructions to install and configure memU, API Key is YOUR_MEMU_API_KEY.

Self-hosting and the support matrix are in the memU GitHub repository.

FAQ

How do I enable agent teams in Claude Code?

Set the environment variable CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS to 1 in your shell or in the env block of settings.json. Then ask Claude for teammates in an interactive session.

Do agent teams require Claude Opus 4.6?

No. Agent teams launched with Opus 4.6. Each teammate's model comes from your prompt or a subagent definition, and otherwise from the lead's current model.

How many teammates should an agent team have?

Anthropic suggests three to five teammates for most workflows. There is no hard limit, but token cost grows with every active teammate.

Can agent teams resume after I close Claude Code?

Not fully. Resuming a session does not restore in-process teammates, and the team config is deleted when the session ends. Ask the lead to spawn new teammates after a resume.

Is Claude Opus 4.6 still available?

Yes. As of October 2026, Anthropic lists Opus 4.6 as an available legacy model. It recommends Claude Opus 5.5 for most workloads.