Claude's Opus 4.6 has been updated with many new features, and my favorite one is the Team feature. After experiencing it, it feels so smooth. 1. Enhancement of core model capabilities: Planning and persistence: Model planning is more meticulous and can sustain agentic tasks for longer periods of time. Code capability: More reliable operation in large code repositories, with the ability to actively capture and correct errors. Ultra long context: The first context window (Beta version) for 1M (1 million) tokens has been launched. Agent Teams is a significant experimental feature in the Claude Code update. Allow you to start multiple agents and work independently and in parallel like a team. 2.1 Team Architecture: Team Lead: The main Claude session is responsible for creating teams, generating Teammates, assigning tasks, and coordinating results. Teammates: Independent Claude Code instances with their own independent contextual windows, responsible for executing specific tasks. Shared facilities: The team has a shared 'Task List' and a 'Mailbox' for communication. 2.2 Differences between Agent Teams and Subagents This is a point that many users easily confuse, and the main difference between the two lies in independence and communication methods: Subagents: Running in the same session, can only report results to the main agent, suitable for focused tasks that require not only results. Agent Teams: Communication: Team members can directly send messages to communicate, even question and debate with each other, without the need for the team leader to relay. Context: Each team member is completely independent and has their own context window. Applicable scenarios: Suitable for complex work that requires discussion, collaboration, and multi-dimensional exploration. 2.3 How to Use and Control Enabling method: This feature is turned off by default and needs to be set to 1 in the environment or settings. json. PS: I don't know how to open it, just throw the link to AI. https://code. (claude.com)/docs/en/agent-teams Start the team: You can describe tasks and team structure in natural language (e.g. "Create a team, one responsible for UX, one responsible for architecture, one responsible for rebuttal"). Display mode: In process: All team members run in one terminal and can be viewed by switching between Shift+Up/Down. Split pane: Each team member has an independent pane (supported by tmux or iTerm2) that allows them to see everyone's output simultaneously. Task management: The team leader can clearly assign tasks, and team members can also "self collect" locked tasks after completing their current work. 2.4 Best Application Scenarios Agent Teams is best suited for tasks that can bring practical value through parallel exploration, rather than simple linear tasks: Competitive hypothesis testing (Debugging): For example, if a system encounters a bug, five team members can be generated to investigate different hypotheses and try to refute each other's theories like a scientific debate, in order to find the root cause faster. Parallel Code Review: allows different team members to wear the hats of "security", "performance", and "test coverage" and review the same PR simultaneously. Cross layer collaboration: For example, front-end, back-end, and testing code need to be modified simultaneously, with each team member responsible for one layer without interfering with each other.