What are agent teams?
Agent teams are groups of AI agent instances that work on one project in parallel, each with its own context window, coordinating through a shared task list, direct messages or file locks.
One public demonstration is Anthropic's "Building a C compiler with a team of parallel Claudes" (February 2026). Sixteen agents ran across nearly 2,000 Claude Code sessions over about two weeks, for just under $20,000 in API costs, and produced a 100,000-line compiler in Rust that can build Linux 6.9 on x86, ARM and RISC-V. Agents claimed work by writing lock files into a shared git repository.
Claude Code now ships agent teams as a feature, experimental and disabled by default. One session acts as the team lead, teammates work in their own context windows, and they share a task list and message each other directly.
Agent teams compared with subagents
A subagent is a helper: it gets a task, works in its own context and reports a result back to the agent that called it. Teammates are peers: each is a full, independent session that can claim tasks, share findings and challenge another teammate's conclusions without going through the lead.
That independence costs tokens. Claude Code's documentation says teams use significantly more tokens than a single session, scaling with the number of active teammates, and points to subagents or a single session for sequential work, same-file edits or tasks with many dependencies.
What makes a team work
A near-perfect oracle
In the compiler project the test suite decided what was correct. Without a strong verifier, agents solve the wrong problem in parallel.
Splittable problems
Parallel agents stalled on one monolithic task, compiling the kernel, until GCC was used as a reference to split it into many independent bugs.
Separate ownership
Two teammates editing the same file overwrite each other. Work is split so each owns its own files.
Right-sized tasks
Claude Code's docs suggest starting with three to five teammates and about five or six self-contained tasks each.
Specialist roles
The compiler team included agents for removing duplicate code, performance, output quality, design critique and documentation.
Quality gates
Hooks can block a task from being marked complete until checks pass.
Where teams struggle
Google Research's study of 180 agent configurations (January 2026) found multi-agent coordination cut performance by 39% to 70% on sequential reasoning, and that agents working without cross-checks amplified errors. Claude Code's own list of limitations includes task status that can lag and teammates that are not restored when a session is resumed. Teams fit broad, separable work; they are a poor fit for a single chain of dependent steps.
How Motiontic's engineers apply it
Motiontic's engineers use parallel agents for separable work with a strong test suite behind it, such as migrations across many modules, and start with research and review tasks before parallel implementation. The test suite is built or hardened first, because it is the only thing keeping a team of agents pointed at the same goal.
Subagents compared with agent teams, per Claude Code's documentation
| Subagents | Agent teams | |
|---|---|---|
| Context | Own window; result returns to the caller | Own window; fully independent |
| Communication | Report back to the calling agent | Teammates message each other directly |
| Coordination | The main agent manages all work | Shared task list and self-claiming |
| Token cost | Lower; results are summarised back | Higher; each teammate is a full instance |
| Fits | Focused tasks where only the result matters | Work that needs discussion and parallel ownership |
Common questions
What is the difference between agent teams and a multi-agent system?
Agent teams are one kind of multi-agent system: peers that coordinate through shared tasks and messages. Other kinds, such as orchestrator and workers, route all coordination through a lead agent.
How many agents should be on a team?
Claude Code's documentation suggests three to five teammates for most workflows, noting that token costs scale linearly and that three focused teammates often outperform five scattered ones.
Are agent teams ready for production work?
In Claude Code they are labelled experimental, with known limits around resuming sessions, task status and shutdown. They suit supervised engineering work more than unattended production systems today.
Related
Subagent
A helper agent with a clean context that hands back a short answer.
Orchestrator-worker pattern
A lead agent plans, workers run in parallel with their own context, the lead merges.
Single Agent vs Multi-Agent
One loop with good tools, compared with several agents splitting the work.
Claude Code vs Cursor
Anthropic's coding agent and the SpaceX-owned AI editor compared on plans, models and agent tooling.
AI agent development
Agents that do real work for users, with the evals, traces and guardrails to keep them in production.
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.