Agent Teams
Agent Teams is a multi-agent collaborative architecture pattern composed of multiple AI agents, each maintaining an independent context window, organized into a networked cooperative structure. Individual agents can process subtasks in parallel and coordinate through inter-agent communication to achieve task division, collectively completing multi-step and multi-dimensional tasks that a single agent cannot handle. Core characteristics include parallel task decomposition, independent context isolation, and dynamic collaboration with information exchange among agents.
Core Facts
Timeline (last 90 days)
Agent Teams consume a large number of tokens; a single sub-agent is sufficient for most tasks, making Agent Teams overkill.
QA scenarios are better suited to be addressed with Skills rather than an entire Agent Team, because a QA process is essentially step-by-step in nature.
The team lead can detect when a teammate process has exited and will sometimes proactively remove teammates that are too idle.
Teammates are created instantly; users simply tell Claude Code in one sentence, and it will automatically generate a team and configure the corresponding system prompts.
In older versions of Cline, every streaming output chunk and every heartbeat at 2-second intervals would trigger a complete re-save of the entire team state
The version used in the cc-haha demo is 0.6.3, and the Agent Teams configuration item is enabled by default
Agent Teams is an experimental feature officially proposed by Anthropic for the first time
There is a communication protocol among Agents in Agent Teams, where the team leader issues assignments and direct messages to members, and members report progress back to the team leader
The theoretical foundation of Agent Teams is Multi-Agent System (MAS)
Anthropic has built Agent Teams directly into the SDK layer of Claude Code, eliminating the cost for users to manually set up an orchestration layer
10 more timeline events
All Facts (20)
The project adopts a multi-agent collaborative architecture, decomposing job-seeking tasks into independent agents with different responsibilities.
75%VerifiedAgent Teams support multiple roles running in parallel sessions.
75%Unverified代理团队模式采用网状拓扑结构，成员间可以点对点沟通，每个成员拥有独立的200K Token上下文窗口
90%Unverified子Agent的生命周期很短，只在执行任务过程中存在，任务完成后就被销毁；而Agent Teams中的每个Agent都是长期存在、持久性的
90%Unverified选择子代理还是代理团队的核心判断标准是任务的独立性与协作需求
85%Unverified代理团队的核心价值在于能够交付单个AI永远无法达到的完整度和专业度，而非速度优势
80%UnverifiedAgent Teams consume a large number of tokens; a single sub-agent is sufficient for most tasks, making Agent Teams overkill.
50%UnverifiedTeammates are created instantly; users simply tell Claude Code in one sentence, and it will automatically generate a team and configure the corresponding system prompts.
50%UnverifiedThe team lead can detect when a teammate process has exited and will sometimes proactively remove teammates that are too idle.
50%UnverifiedQA scenarios are better suited to be addressed with Skills rather than an entire Agent Team, because a QA process is essentially step-by-step in nature.
50%UnverifiedIn older versions of Cline, every streaming output chunk and every heartbeat at 2-second intervals would trigger a complete re-save of the entire team state
50%UnverifiedAgent Teams is an experimental feature officially proposed by Anthropic for the first time
50%UnverifiedThere is a communication protocol among Agents in Agent Teams, where the team leader issues assignments and direct messages to members, and members report progress back to the team leader
50%UnverifiedThe theoretical foundation of Agent Teams is Multi-Agent System (MAS)
50%UnverifiedAnthropic has built Agent Teams directly into the SDK layer of Claude Code, eliminating the cost for users to manually set up an orchestration layer
50%UnverifiedUsing Agent Teams in cc-haha only requires updating it to the latest version, confirming the configuration item is enabled, and adding 'use agent teams to have them collaborate' to the prompt
50%UnverifiedThe SDK layer code for Agent Teams and the visualization code for cc-haha are both publicly available in open-source repositories
50%UnverifiedIn multi-agent parallel scenarios, the more shared content there is and the more parallel rounds there are, the more significant the caching benefits
50%UnverifiedThe Tongyi Agent Teams system consists of four agents — screenwriter, director, art director, and video generation — allowing users to input a story concept and have tasks automatically divided from script expansion to final rendering.
50%UnverifiedSubagent的隔离策略牺牲协作深度换取信息纯净度，Agent Teams的共享策略接受一定上下文复杂性以换取实时协同能力
50%