Mail

For a long time, even AI agents were stuck in a "working alone" structure. Claude Code's Agent Teams feature, released in June 2026, breaks through that limitation.

Multiple Claude Code sessions form a team. One becomes the Team Lead, the rest are Teammates. The lead coordinates and distributes work, while teammates operate independently in their own context windows — and can communicate directly with each other.

What Are Agent Teams?
How the Team Lead / Teammate Structure Works
How It Differs from Nested Sub-Agents
The Most Effective Use Cases
Cost and Token Strategy
How to Enable It
Implications for EdTech

1. What Are Agent Teams?

Multiple AI sessions communicate with each other to complete a single project.

Traditional Claude Code used a vertical hierarchy — sub-agents spawned inside one session. Agent Teams goes further: fully independent Claude Code sessions collaborate horizontally.

The team lead session understands the whole project and designs the work. Teammate sessions each handle their assigned task inside their own full context window and report results back to the lead.

Claude Code Agent Teams Structure

Agent Teams is currently an experimental feature, disabled by default. It is part of the CLI 2026.6 release train.

2. How the Team Lead / Teammate Structure Works

Role	Responsibilities
Team Lead	Plans work, assigns tasks, synthesizes results
Teammates	Execute tasks independently, communicate directly with lead and each other

Direct teammate-to-teammate communication is the key innovation. In traditional multi-agent systems, all information had to pass through a central coordinator. With Agent Teams, teammates can share context directly.

3. How It Differs from Nested Sub-Agents

Claude Code launched Nested Sub-agents earlier in June — it is easy to confuse the two.

Feature	Nested Sub-Agents	Agent Teams
Execution	Hierarchical within one session	Cooperative independent sessions
Context	Parent–child structure	Each session has its own context
Communication	Parent → child only	Direct peer communication
Depth limit	Up to 5 levels	Scales by number of teammates
Token cost	Compounds per level	Scales per session

Simply put: Nested Sub-Agents = one agent playing multiple roles. Agent Teams = multiple people actually working together.

4. The Most Effective Use Cases

① Research & Code Review Teammates simultaneously analyze different modules — one for security, another for performance, another for test coverage. The lead synthesizes the full review.

② New Feature Development Frontend, backend, and tests developed in parallel by separate teammates. Teammates negotiate API contracts directly, reducing integration conflicts.

③ Multi-Hypothesis Debugging Multiple teammates test different theories about the same bug simultaneously. The lead adopts the result from whichever teammate finds the root cause first.

④ Cross-Layer Refactoring Large refactors touching frontend, backend, and infrastructure handled by dedicated teammates for each layer — all running in parallel.

5. Cost and Token Strategy

Each teammate session has its own full context window. Three teammates = roughly 3× the token cost. Add the lead's context on top and costs can be higher.

Cost-reduction tips:

Break tasks small enough that each teammate context stays lean
Use only as many teammates as needed (3–5 fits most cases)
Assign cheaper models to teammate sessions
Keep batchable, sequential work as a single agent

6. How to Enable It

# Add to settings.json or environment
CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1

Configure teams with the /team command or an agent-teams settings block. Requires CLI version 2026.6 or later.

Since this is experimental, test in development or staging before using in production.

7. Implications for EdTech

For those building EdTech products, Agent Teams is interesting for more than raw development speed.

Imagine a learning content production pipeline where one agent analyzes the curriculum, another generates practice problems, and a third validates difficulty — all in parallel. Work that today takes weeks of manual effort could compress into hours.

As AI agents evolve from tools into teammates, studying how humans and AI can best collaborate becomes the more important question.

Wrap-Up

Agent Teams is experimental, but the direction is clear. AI agents collaborating like a human team is no longer a distant vision — you can try it today with one line in settings.json.

Further Reading

Sources:

Claude Code Agent Teams: Multiple AI Sessions Working Together as a Team

Table of Contents