Building Agent Teams with the Anthropic SDK
What if building an AI agent team was as easy as answering a few questions?
Most agent frameworks demand upfront configuration. Attune guides you through decisions instead. We call this Socratic discovery.
Step 1: Start with /attune
Type /attune in Claude Code. The framework asks what you need:
? What are you trying to accomplish?
○ Fix or improve something
○ Build or create something
○ Test or validate something
○ Review or analyze something ← you pick
No commands to memorize. Just answer.
Step 2: Scope Your Work
? Which area should we focus on?
○ Specific files
○ Recent changes
○ Full project ← you pick
? What kind of review?
○ Security audit
○ Code quality
○ All of the above ← you pick
The MetaOrchestrator scores its confidence. Clear tasks (>80%) run automatically. Vague requests trigger more questions.
Step 3: The Framework Proposes a Team
? How would you like to create the agent team?
○ Use recommended: parallel (Recommended)
3 agents: Security Expert, Code Reviewer,
Documentation Checker. Confidence: 85%
○ Customize team composition
○ Show all patterns
Accept or customize. Customization gives you a multi-select agent picker and pattern selector — still guided.
Step 4: Agents Run with Cost Control
Every agent starts on the cheapest tier (Haiku). If it can't handle the task, it escalates to Sonnet, then Opus. You only pay for what you need.
✓ Running review with 3 agents...
Security Auditor: score 92 (cheap, $0.003)
Code Reviewer: score 85 (cheap, $0.003)
Doc Checker: score 45 (cheap, $0.003)
Quality Gates:
security_score (≥80): ✓ PASS
quality_score (≥70): ✓ PASS
doc_coverage (≥60): ✗ FAIL (advisory)
Team Result: SUCCESS — total $0.009
Required gates block on failure. Advisory gates report but don't block.
What Just Happened?
Three questions in, you have a parallel agent team with quality enforcement — no boilerplate, no config. The Socratic approach means:
- Beginners get guided through every decision
- Experts get automatic execution when confidence is high
- Everyone gets scoped, cost-efficient runs
Go deeper with SDKAgent and SDKAgentTeam in code, or explore all 10 composition patterns in the Grammar of AI Collaboration series.
pip install attune-ai[developer]
python -m attune.models.auth_cli setup
Type /attune and let the questions guide you.
Attune AI is open-source. Star us on GitHub.
Related Articles
How to Build AI Agents with Claude: A Step-by-Step Tutorial
A practical, code-first tutorial for building AI agents with Claude and Attune AI. Covers single agents, multi-agent teams, tool use, and state persistence.
Prompt Caching with Anthropic: Save 90% on Claude API Costs
Anthropic's prompt caching can reduce your Claude API costs by up to 90%. Here's how it works, when to use it, and how Attune AI enables it automatically.
Multi-Agent Orchestration Patterns for AI Developers
Six proven multi-agent orchestration patterns with Python code examples: parallel, sequential, delegation, two-phase, quality-gated, and escalation chains.