An agent team is a group of artificial intelligence (AI) agents that work on one goal at the same time, each in its own session, and coordinate with each other rather than only with the person who started them. One agent usually acts as the lead. It breaks the goal into tasks, starts the other agents, known as teammates, and assembles the result. Teammates pick up tasks and can message one another directly. The arrangement exists because a single large language model (LLM) agent has one context window, one chain of reasoning and one set of permissions, and broad or parallel work, such as reviewing a large codebase from several angles or auditing a campaign set-up against different rule books, fits poorly into that single thread.

The term has no formal standard. Anthropic uses "agent teams" as a product name in Claude Code, Microsoft's AutoGen uses "team" as a class of objects, and CrewAI calls the same idea a "crew". Regulators now use it as well. The UK Information Commissioner's Office (ICO), in a call for evidence dated October 8, 2026, separates subagents, "which work for an orchestrating agent and may share its session," from agent teams, which "have separate contexts, tools or permissions and can talk to each other and to the user," according to PPC Land's report on the consultation.

How an agent team works

Anthropic's documentation for Claude Code gives the most detailed public description of the mechanics. A team has four parts. The team lead is the main session that spawns teammates and coordinates work. Teammates are separate Claude Code instances, each with its own context window. The task list is a shared list of work items that teammates claim and complete. The mailbox is the messaging system that carries notes between agents.

The steps run roughly as follows.

  1. A person asks for a team, or the lead proposes one. The lead writes a team configuration with a members array, its own entry typed team-lead.
  2. The lead creates tasks. Each task is pending, in progress or completed, and a task can depend on others. A task with unfinished dependencies shows as blocked, and teammates do not claim it.
  3. Teammates claim tasks, either by assignment from the lead or by taking the next unblocked item themselves. Two agents can try to claim the same task at once, so claiming uses file locking to prevent race conditions, according to Anthropic.
  4. Agents exchange messages. In Claude Code each agent has an inbox stored as a JSON file at ~/.claude/teams/{team-name}/inboxes/{agent-name}.json, and a message counts as sent only when that write succeeds.
  5. Blocked tasks unblock when their dependencies complete. The lead collects findings and reports back.

Hooks named TeammateIdle, TaskCreated and TaskCompleted let an operator run a script at each point. A hook exiting with code 2 can keep an idle teammate working or block a task. A teammate spawned while the lead is in plan mode works read-only until its plan is approved. Teammates run inside one terminal or in split panes through tmux or iTerm2, set by teammateMode.

Anthropic's guidance on size is modest: "Start with 3-5 teammates for most workflows," with five to six tasks per teammate. There is no hard limit. As of November 2026 the feature remains labelled experimental and disabled by default, switched on with the environment variable CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS set to 1. Listed limitations include one team per session, no nested teams, a fixed lead and task status that "can lag".

Other team designs

Frameworks implement the idea differently, mainly in how they decide who acts next. Microsoft's AutoGen AgentChat defines a team as "a group of agents that work together to achieve a common goal" and ships preset classes. RoundRobinGroupChat gives agents turns in a fixed order. SelectorGroupChat asks a model to choose the next speaker after each message. Swarm passes control through a HandoffMessage. A team stops when a termination condition fires, for example when an agent writes the word "APPROVE".

CrewAI's crews run either a sequential process, where tasks execute one after another, or a hierarchical one, where "a manager agent coordinates the crew, delegating tasks and validating outcomes before proceeding," according to its documentation.

Some teams have no lead at all. Nicholas Carlini, a researcher on Anthropic's Safeguards team, ran 16 Claude agents in parallel to write a C compiler, published on February 5, 2026. Each agent worked against a shared git repository and claimed a task by writing a text file into a current_tasks/ folder. "I don't use an orchestration agent," Carlini wrote. Over nearly 2,000 sessions and about $20,000 in API costs, the agents produced a 100,000-line compiler able to build Linux 6.9 on x86, ARM and RISC-V.

Origin and evolution

Multi-agent systems are an old research field, but LLM agent teams date from August 2023. MetaGPT, submitted on August 1, 2023, with Sirui Hong as lead author, encoded standard operating procedures into prompts and gave agents assembly-line roles such as product manager and engineer. AutoGen, submitted on August 16, 2023, with Qingyun Wu as lead author, described applications built from "multiple agents that can converse with each other to accomplish tasks."

Production systems followed in 2025. OpenAI released its Agents SDK on March 11, 2025, replacing its experimental Swarm library, which reduced coordination to agents and handoffs. On June 13, 2025, Anthropic described its Research feature, in which a Claude Opus 4 lead agent directed Claude Sonnet 4 subagents and outperformed a single Opus 4 agent by 90.2% on an internal research evaluation. Workers reported only to the lead.

The shift to peer communication came with Claude Opus 4.6 on February 5, 2026, when Anthropic introduced agent teams in Claude Code as a research preview. Agents "work in parallel as a team and coordinate autonomously," the company said, adding that the feature was "best for tasks that split into independent, read-heavy work like codebase reviews."

Why it matters for marketing

Advertising work increasingly runs as multi-step agent workflows, and the design question is whether to give one agent everything or split the work. Meta answered it in October 2026 by splitting one AI agent into 14 specialists after it hallucinated. A 1,300-line orchestrator now routes more than 20 data-engineering workflows to subagents with isolated context, and the orchestrator itself "performs zero domain work," according to Meta's engineers. Configuration for a new data column now takes under an hour, down from three or four days.

Anthropic's own marketing operations team uses a similar split. Its event-build workflow runs a dispatcher skill that does no set-up itself plus five specialist skills, and an audit step runs on a fresh Claude instance with no prior context, which submits a test registration and checks the confirmation email.

Separate contexts and permissions also make agent teams harder to trace than a single assistant. The ICO warned that orchestration can let "inaccurate information, inferences or bias to spread quickly through the data-processing chain," and its call for evidence on agentic AI closes on November 20, 2026.

Limitations and disputes

Cost is the first objection. Anthropic found that agents typically use about four times more tokens than chat interactions and multi-agent systems about 15 times more. Token use alone explained 80% of performance variance in Anthropic's BrowseComp tests, so part of a team's edge is simply spending more.

Reliability is the second. Researchers led by Mert Cemri at UC Berkeley, in a paper first submitted on March 17, 2025, built a taxonomy of 14 failure modes in three categories: system design, inter-agent misalignment and task verification. Drawing on more than 1,600 traces across seven frameworks, it found that "performance gains on popular benchmarks are often minimal." A SUSTech preprint posted on October 30, 2026, cited the same data, with 44.2% of labelled failures in system design and 32.3% in inter-agent misalignment, and argued that checking each agent cannot replace checking the system.

The sharpest dispute is architectural. Walden Yan of Cognition published "Don't Build Multi-Agents" on June 12, 2025, arguing that "actions carry implicit decisions, and conflicting decisions carry bad results." Anthropic conceded part of the point the next day, writing that "most coding tasks involve fewer truly parallelizable tasks than research." Carlini's compiler shows the residue: merge conflicts were "frequent", and "it is easy to see tests pass and assume the job is done."

Oversight has gaps of its own. In Claude Code, a teammate's plan can be approved automatically "without the lead reviewing it" if its permission mode was never changed, according to Anthropic's documentation.

Agent team versus adjacent terms

Subagent. A subagent is a helper spawned for one focused task. It has its own context window but returns a result only to the agent that called it. Teammates are fully independent sessions that message each other and share a task list.

Multi-agent system. The umbrella term for any software with more than one agent. An agent team is one kind, defined by peer messaging and shared work tracking.

Orchestration layer. An orchestration layer is the coordinating software, which can be deterministic code. In a team, coordination is partly done by the agents themselves.

Multiplayer AI. Multiplayer AI means several people sharing one assistant, as in Yext's Scout workspace. An agent team means several agents serving one goal, often for one person.

Recent developments

Claude Code agent teams remain experimental as of November 2026, yet the ICO already treats agent teams as a category distinct from subagents. Meta's thin-orchestrator post, covered by PPC Land on October 19, 2026, and the SUSTech preprint covered on November 3, 2026, show the debate shifting from whether to split agents to how much reasoning the coordinator should do. In Meta's design, only one of seven protocol steps uses LLM reasoning at all.

Timeline

  • August 1, 2023 - MetaGPT paper submitted, assigning software-company roles to LLM agents.
  • August 16, 2023 - AutoGen paper submitted, describing conversing multi-agent applications.
  • March 11, 2025 - OpenAI releases the Agents SDK, succeeding the experimental Swarm library.
  • March 17, 2025 - UC Berkeley researchers submit "Why Do Multi-Agent LLM Systems Fail?" with 14 failure modes.
  • June 12, 2025 - Cognition's Walden Yan publishes "Don't Build Multi-Agents".
  • June 13, 2025 - Anthropic details its multi-agent Research system and the 90.2% result.
  • February 5, 2026 - Anthropic launches Claude Opus 4.6 and agent teams in Claude Code as a research preview.
  • February 5, 2026 - Nicholas Carlini publishes the 16-agent C compiler project.
  • July 8, 2026 - Anthropic marketing operations describes dispatcher and audit agents in Cowork.
  • October 8, 2026 - The ICO dates its agentic AI call for evidence defining agent teams.
  • October 19, 2026 - PPC Land reports Meta's 14-specialist thin-orchestrator design.
  • October 30, 2026 - SUSTech preprint argues no algorithm can certify agent goal retention.
  • November 20, 2026 - ICO call for evidence closes.

Summary

Who: Agent teams are built by AI developers and platform operators, with Anthropic (Claude Code agent teams), Microsoft (AutoGen teams) and CrewAI (crews) offering frameworks, and companies such as Meta and Anthropic's own marketing team running split-agent workflows; regulators including the UK ICO now define the term.

What: A group of AI agents, each with its own context, tools or permissions, that split one goal into tasks, claim work from a shared list and message each other, usually under a lead agent.

When: LLM agent teams emerged in research in August 2023, reached production frameworks in 2025, and became a named Claude Code feature on February 5, 2026; it remains experimental as of November 2026.

Where: In coding tools, AI agent frameworks, data-engineering pipelines and marketing operations workflows, running on developer machines, containers or cloud platforms.

Why: A single agent's context window and reasoning thread limit broad or parallel work; teams spread that work across independent sessions, at the price of far higher token use, coordination failures and harder oversight.