AI agents · OpenClaw · self-hosting · automation

Quick Answer

How to Use Claude Agent Teams (Opus 4.6 Guide)

Published: • Updated:

How to Use Claude Agent Teams (Opus 4.6 Guide)

Claude Opus 4.6’s agent teams let you split larger tasks across multiple coordinated AI agents that work in parallel. Instead of one agent working through tasks sequentially, agent teams divide work and coordinate directly—completing projects faster like a talented human team.

What Are Agent Teams?

Released February 5, 2026, agent teams are Anthropic’s approach to multi-agent orchestration:

“Instead of one agent working through tasks sequentially, you can split the work across multiple agents — each owning its piece and coordinating directly with the others.” — Anthropic

Key benefits:

  • Parallel execution - Multiple agents work simultaneously
  • Coordination - Agents communicate and hand off work
  • Specialization - Each agent can own a specific domain
  • Faster completion - Like having a skilled team vs one person

How Agent Teams Work

  1. Task decomposition - The primary agent breaks a project into sub-tasks
  2. Agent spawning - Sub-agents are created for each task
  3. Parallel work - Agents execute simultaneously
  4. Coordination - Agents share results and dependencies
  5. Assembly - Results combine into final output

Getting Started

Requirements

  • Claude API access (Anthropic account)
  • Opus 4.6 model access
  • Agent teams in research preview

Basic Setup

from anthropic import Anthropic

client = Anthropic()

# Create a task that benefits from agent teams
response = client.messages.create(
    model="claude-opus-4-6-20260205",
    max_tokens=8192,
    system="""You are an AI assistant with agent teams capability.
    When given complex tasks, split them across specialized agents:
    - Research agent: Gathers information
    - Analysis agent: Processes findings
    - Writing agent: Creates output
    Coordinate agents to work in parallel.""",
    messages=[{
        "role": "user",
        "content": "Research the top 5 AI frameworks, analyze their strengths/weaknesses, and write a comparison report."
    }]
)

Using Adaptive Thinking

Opus 4.6 recommends adaptive thinking for agent teams:

response = client.messages.create(
    model="claude-opus-4-6-20260205",
    max_tokens=8192,
    thinking={
        "type": "adaptive"  # Recommended over manual budget
    },
    messages=[...]
)

Note: Manual thinking (type: "enabled" with budget_tokens) is deprecated. Use adaptive thinking.

Example Use Cases

1. Codebase Analysis

Single agent approach: Sequentially analyzes each file, then patterns, then suggestions.

Agent teams approach:

  • Agent A: Analyzes authentication code
  • Agent B: Analyzes database layer
  • Agent C: Analyzes API endpoints
  • Agent D: Reviews test coverage
  • Coordinator: Synthesizes findings

Result: 4x faster analysis with specialized attention to each area.

2. Research Report

Agent teams breakdown:

  • Research agents (3): Each researches different sources
  • Fact-checker agent: Validates claims
  • Writing agent: Synthesizes into report
  • Editor agent: Reviews and polishes

3. Multi-File Refactoring

Agent teams breakdown:

  • Planning agent: Creates refactoring strategy
  • Refactor agents (N): Each handles a module
  • Test agent: Updates and runs tests
  • Integration agent: Ensures compatibility

Best Practices

1. Clear Task Decomposition

Break tasks into independent sub-tasks that can run in parallel.

Good: "Research X, Research Y, Research Z, then combine"
Bad: "Research X, then use X results to research Y"

2. Define Agent Specializations

Give each agent a clear role and domain.

3. Plan Coordination Points

Identify where agents need to share information or wait for others.

4. Use for Complex Tasks

Agent teams add overhead. Use them for tasks that genuinely benefit from parallelization.

Good use cases:

  • Multi-file code analysis
  • Research across multiple sources
  • Large document processing
  • Comprehensive audits

Less suited:

  • Simple Q&A
  • Single-file edits
  • Linear workflows

Pricing Considerations

Agent teams use more tokens (multiple agents = multiple conversations), but complete faster.

Opus 4.6 pricing:

  • Input: $5/M tokens
  • Output: $25/M tokens
  • Caching: Up to 90% savings

Tip: Enable prompt caching to reduce costs when agents share context.

Current Status (March 2026)

  • Availability: Research preview
  • Access: API users, Max/Team subscribers
  • Model: Claude Opus 4.6 required

Last verified: March 11, 2026