What are agent teams in AI coding?
By Flavio
Agent teams coordinate several coding agents through shared tasks and messages. Learn how they differ from subagents and when parallel work helps.
Agent teams are groups of coding agents that coordinate work through assignments, messages, and a shared goal. A lead divides the work and brings the results together.
The name can mean a general workflow or a particular product feature. Claude Code has a feature called agent teams. Other tools expose parallel work through subagents. For the practical coordination side, read Parallel coding agents: how to coordinate and review changes.
How agent teams work
Each worker has its own conversation context. Someone needs to give it the requirements, decide which files it owns, and review what comes back. Starting several sessions also creates several sets of assumptions to reconcile.
Agent teams in each tool
Claude Code
Agent teams are experimental and disabled by default. Enable them with CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1. A lead coordinates independent teammates, which can message each other and share tasks. Teammate spawning requires an interactive session; the -p mode runs named helpers as ordinary subagents instead.
Use teams when workers need to discuss findings. A focused job that only needs a result back can use a subagent. Teams add coordination overhead and token usage.
OpenAI Codex
Codex documents subagents for splitting work into specialized threads. Give each worker a bounded task and let the parent combine the results. Use the documented Codex configuration rather than copying Claude Code’s team setting.
Cursor
Cursor’s subagents have separate context windows and return results to the parent. They can work in parallel. By default they share the parent’s checkout, so simultaneous edits need clear file ownership.
Grok Build
Grok Build’s subagents are independent child sessions that return summaries to the parent. Its built-in types include general-purpose, explore, and plan. The latter two have no shell or edit tools. The page does not document a shared team task board.
Details checked against each vendor’s documentation in November 2026.
A real example
For an interactive Claude Code session with teams enabled, this prompt gives the workers separate review jobs:
Create a team to review this repository's AI glossary.
One teammate checks internal links. Another checks publication dates.
Both should report findings without editing files.
Compare the findings and return file paths and the fixes needed.
Use parallel review when the checks are independent. Keep a change in one session when each step depends on the previous result, or when everyone would edit the same file.
Pitfalls
Separate conversations can disagree about the same requirement. Put shared decisions in project documentation and include them in each assignment. Review the combined result even when every worker reports success.
Separate context windows also do not separate files. Check the tool’s workspace behavior before asking several workers to write. Git worktrees explain one way to isolate changes.
For following several sessions, Herdr covers terminal workspaces. Architecture Dissector helps you inspect the application those sessions are changing; the architecture tutorial walks through it.
Related glossary entries
- Subagents: A subagent is a helper your coding agent starts with its own context window. How subagents work in Claude Code, Codex, Cursor and Grok Build.
- Git worktrees: A git worktree is a second checkout of one repo, so an agent can edit while your checkout stays put. How Claude Code, Codex, Cursor and Grok Build start one.
- AI code review: AI code review is an agent reading a diff for bugs. How Claude Code, Codex, Cursor Bugbot and Grok Build run that review.
Want me to talk about your product? You can sponsor this site.