# Claude Code’s Agent Teams, Nine Months Later

> Source: <https://pub.towardsai.net/claude-codes-agent-teams-nine-months-later-20f28fd1abbd?source=rss----98111c9905da---4>
> Published: 2026-11-20 14:31:01+00:00

**Key takeaway:** Agent Teams requires setting CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1, and unlike subagents, each teammate keeps its own context window for the team's life and messages peers directly, so token usage scales with the number of concurrent context windows.

Open a terminal, ask Claude Code to review a pull request, and tell it to spin up three teammates instead of reviewing it alone. One checks for security issues, one checks performance, one checks test coverage. None of them reports back to a single coordinator and waits its turn. Each claims a task off a shared list, works at the same time as the others, and messages them directly while you watch. That feature is called Agent Teams, and it shipped as a research preview alongside Claude Opus 4.6 on February 5, 2026 [1]. Nine months on, it’s worth asking what actually stuck.

**In this article:**

Agent Teams is not a model capability. It’s a Claude Code orchestration feature, and it stays off unless you set CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 in your settings or shell environment. Once it's on, one Claude Code session becomes the team lead: it spawns teammates, assigns work, and synthesizes what comes back. Each teammate is a separate Claude Code instance with its own context window, not a lightweight call that returns a string to the caller [2].

Coordination runs through a task list that holds work items in three states (pending, in progress, completed). Teammates either get assigned a task by the lead or self-claim the next open one, with file locking so two teammates can’t grab the same task at once [2].

Underneath that sits a mailbox for each agent: a JSON file at a fixed path under ~/.claude/teams/, where messages from other agents land and get validated on read. One engineer's working example, split across a frontend teammate, an API teammate, and a migration teammate, had each one "own its piece" and coordinate directly with the others rather than routing everything through a lead [3]. That's a more literal architecture than "the agents talk to each other" suggests. It behaves closer to a small message queue than a conversation.

Claude Code already had subagents before Agent Teams existed, and the two are easy to conflate because both parallelize work. The distinction is about where control and context live. A subagent gets a task, does it in its own context window, and returns a summarized result to the main agent, which stays in charge of the whole job. A teammate keeps its context window for the life of the team, talks to other teammates without going through the lead first, and can be addressed by you directly, by name, at any point [2].

That difference is exactly the one the Claude Agent SDK’s coordinator and subagent pattern is built around: a coordinator that owns the plan, and subagents that execute pieces of it and hand results back. If you’ve read [how that pattern works in the SDK](https://medium.com/gitconnected/coordinator-subagent-orchestration-in-claude-agent-sdk-a127eacc6f5b), Agent Teams is what happens when you remove the “hand results back to one owner” constraint and let the workers negotiate among themselves instead. The tradeoff is explicit in Anthropic’s own comparison table: subagents cost less because results get summarized back into one context, while a team’s token usage scales with however many independent context windows are running at once [2].

Anthropic’s own use-case examples are narrower than “multi-agent coordination” might suggest, and that’s informative on its own. The first is a parallel code review: three teammates each take a lens, security, performance, test coverage, on the same pull request, and the lead merges their findings afterward [2].

The second is adversarial debugging. Several teammates investigate competing hypotheses for the same bug and are explicitly told to try to disprove each other’s theories before converging on a shared findings document. Anthropic frames this around a known failure mode of single-agent debugging: once an agent settles on one plausible explanation, it tends to stop looking for a better one, and running several independent investigators against each other is the proposed fix [2].

Both examples share a property the documentation states directly: the work has to split into pieces that don’t depend on each other, or file conflicts and coordination overhead erase the benefit fast. Anthropic’s own sizing guidance is to start with three to five teammates for most workflows, not to scale up by default [2]. That’s a narrower claim than the launch framing of “multiple agents coordinating autonomously,” and it matters if you were picturing something closer to [a workflow that fans dozens of agents across phases](https://medium.com/towards-artificial-intelligence/workflows-vs-agents-a-decision-framework-852abc08e12f) from one instruction.

Opus 4.6’s launch leaned on several specific figures. On the 8-needle variant of MRCR v2, a long-context retrieval benchmark, Opus 4.6 scored 76% against Sonnet 4.5’s 18.5%. On GDPval-AA, an evaluation Artificial Analysis ran independently, Opus 4.6 beat GPT-5.2 by roughly 144 Elo points, which the footnote translates to winning about 70% of head-to-head comparisons. On BrowseComp, a multi-agent harness pushed the score to 86.8% [1].

Two of the claims are worth being more careful with. Anthropic’s page says Opus 4.6 “achieves the highest score” on Terminal-Bench 2.0, but gives no number in the body text, and I couldn’t find one in Anthropic’s own materials to cite here. Treat that specific claim as unverified rather than a quantified result. The Humanity’s Last Exam score is the more interesting case: Anthropic initially reported 53.1% on the “with tools” setup, then corrected it to 53.0% on February 23, 2026, after an improved cheating-detection pipeline flagged three additional contaminated instances the original pipeline had missed [1]. A correction that small is a good reminder of how much benchmark evaluation still depends on the pipeline doing the scoring, not only on the model being scored.

What’s notable, nine months on, is that Agent Teams hasn’t graduated out of its experimental flag. The current Claude Code documentation opens with the same warning it shipped with: the feature is experimental and off by default, with known limitations around session resumption, task coordination, and shutdown behavior. A session resumed with /resume doesn't bring in-process teammates back, teams don't nest, and a session can only ever have one team [2]. Opus itself has moved on in the meantime: Opus 4.7, 4.8, and 5 all shipped after 4.6, and Opus 5.5 is the current flagship as of this writing [4].

None of that reads as abandonment. The feature still ships, the documentation is still maintained, and the two headline use cases, parallel review and adversarial debugging, are still the ones Anthropic leads with. It reads more like a feature Anthropic is comfortable shipping as permanently experimental: useful enough to keep, risky enough that defaulting it to off still makes sense across four model releases.

The honest framing in Anthropic’s own documentation is a cost tradeoff, not a capability upgrade: a team of teammates is worth it when a task genuinely splits into independent pieces that benefit from being explored at once, and worth avoiding when it doesn’t. For sequential work, same-file edits, or anything with real dependencies between steps, a single session or ordinary subagents are described as the more effective option, in Anthropic’s own wording, not mine [2].

That’s consistent with the economics anyone who has [tracked what an agent loop actually spends on top of model pricing](https://medium.com/towards-artificial-intelligence/why-your-agent-loop-costs-more-than-its-model-price-tag-81a48b691df5) would expect: a teammate is a full session with its own context window, so three teammates reviewing the same pull request cost roughly three sessions’ worth of tokens, not one session split three ways. You can also assign a different model to each teammate, which is its own cost lever worth lining up against [a framework for choosing between Opus, Sonnet, and Haiku](https://medium.com/gitconnected/choosing-between-opus-sonnet-and-haiku-a-cost-framework-7252414a3e1c) before defaulting every teammate to the lead’s model.

Agent Teams never became the default way to run Claude Code, and based on how it’s still documented, it was never meant to. It settled into a specific tool for a specific shape of problem: work you can genuinely split, where several independent perspectives fighting it out beats one agent settling on the first plausible answer.

If you enjoyed the article and wish to show your support, make sure to:

[Claude Code’s Agent Teams, Nine Months Later](https://pub.towardsai.net/claude-codes-agent-teams-nine-months-later-20f28fd1abbd) was originally published in [Towards AI](https://pub.towardsai.net) on Medium, where people are continuing the conversation by highlighting and responding to this story.
