Resources for Agent Orchestration & Safety

Official Documentation

Anthropic

• Claude Code Documentation: https://docs.anthropic.com/en/docs/claude-code
• Claude Code Subagents: https://docs.anthropic.com/en/docs/claude-code/sub-agents
• Claude Code Hooks: https://docs.anthropic.com/en/docs/claude-code/hooks
• Agent SDK: https://docs.anthropic.com/en/docs/build-with-claude/agent-sdk
• Tool Use Guide: https://docs.anthropic.com/en/docs/build-with-claude/tool-use
• Prompt Caching: https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching
• Python SDK: https://github.com/anthropics/anthropic-sdk-python
• TypeScript SDK: https://github.com/anthropics/anthropic-sdk-node

Model Context Protocol (MCP)

• MCP Specification: https://modelcontextprotocol.io/
• MCP Server Registry: https://github.com/modelcontextprotocol/servers
• Building MCP Servers: https://modelcontextprotocol.io/docs/concepts/servers

Orchestration Frameworks

• LangGraph: https://github.com/langchain-ai/langgraph
• LangGraph Documentation: https://langchain-ai.github.io/langgraph/
• CrewAI: https://github.com/crewAIInc/crewAI
• AutoGen: https://github.com/microsoft/autogen

Orchestration Patterns

Framework Comparison

Framework	Pattern	Best For	Complexity
Claude Code subagents	Delegation	Coding tasks, file operations	Low
Custom (Anthropic SDK)	Any pattern	Full control, production systems	Medium
LangGraph	State machines	Complex workflows with branching	High
CrewAI	Role-based teams	Content creation, research	Medium
AutoGen	Conversations	Multi-agent dialogue	Medium

Pattern Reference

# Quick reference for each pattern

# 1. Subagent Delegation
parent_result = synthesise([
    subagent_1.execute(task_a),
    subagent_2.execute(task_b),
])

# 2. Pipeline
result = stage_3(stage_2(stage_1(input)))

# 3. Fan-out / Fan-in
parallel_results = parallel_execute([search_a, search_b, search_c])
result = synthesise(parallel_results)

# 4. LangGraph State Machine
graph.add_edge("plan", "execute")
graph.add_conditional_edges("review", decide, {"revise": "execute", "done": END})

Safety and Cost Management

Safety Checklist

Use this checklist for every production agent system:

• Token budget controller with per-task limits
• Iteration limiter (max loops per agent)
• Timeout controls (max wall-clock time)
• Sandboxed code execution (E2B, Docker, or restricted Python)
• Human-in-the-loop approval for destructive actions
• Audit logging (every tool call, every LLM call)
• Input validation on all tool parameters
• Network restrictions (allowlist for outbound connections)
• Error recovery with retries and fallbacks
• Cost alerting when approaching budget thresholds

Cost Estimation

Model selection has the biggest impact on cost. Check the Anthropic pricing page for current rates. General guidance:

Model	Use For	Relative Cost
Haiku 4.5	Classification, routing, formatting, simple extraction	Lowest
Sonnet 4.6	General work, coding, analysis, synthesis	Medium
Opus 4.8	Complex reasoning, architecture, critical decisions	Highest

Cost reduction strategies:

1. Route simple tasks to Haiku (saves significantly per call)
2. Use prompt caching for repeated system prompts
3. Trim conversation context aggressively
4. Batch multiple classifications into one call
5. Set hard token budgets and stop early

Observability Tools

Tool	Purpose	Link
LangSmith	Agent tracing, debugging, evaluation	https://docs.smith.langchain.com/
Helicone	LLM observability, cost tracking, caching	https://www.helicone.ai/
Braintrust	Evaluation and monitoring	https://www.braintrust.dev/
Weights & Biases	Experiment tracking	https://wandb.ai/

Code Execution Sandboxes

Tool	Type	Link
E2B	Cloud sandboxes	https://e2b.dev/
Modal	Serverless compute	https://modal.com/
Docker	Local containers	Standard Docker
RestrictedPython	In-process sandbox	https://github.com/zopefoundation/RestrictedPython

Research Papers

Multi-Agent Systems

1. ReAct: Synergizing Reasoning and Acting (2023)

   • arXiv: https://arxiv.org/abs/2210.03629
   • Foundation for agent loops

2. Generative Agents: Interactive Simulacra of Human Behavior (2023)

   • arXiv: https://arxiv.org/abs/2304.03442
   • Memory architectures for long-running agents

3. Model Context Protocol Specification (2024-2025)

   • https://modelcontextprotocol.io/
   • Standardised tool integration

Safety and Alignment

4. Constitutional AI (2023)

   • arXiv: https://arxiv.org/abs/2212.08073
   • Self-supervision for safe agent behaviour

5. Agent Safety in Production Systems (2025)

   • Focus: Guardrails, cost control, and audit patterns for deployed agents

Development Setup

Required Packages

# Core
pip install anthropic tenacity python-dotenv

# For LangGraph exercises
pip install langgraph langchain langchain-anthropic

# For testing
pip install pytest

Project Template

my_agent_system/
    orchestrator.py      # Main orchestration
    safety.py            # Budget, limits, audit
    router.py            # Model routing
    agents.py            # Subagent implementations
    config.py            # Configuration
    main.py              # Entry point
    tests/
        test_safety.py
        test_router.py
        test_orchestrator.py
    logs/                # Audit logs (gitignored)
    requirements.txt
    .env                 # API keys (gitignored)
    CLAUDE.md            # Agent configuration

Community

• Anthropic Discord: https://discord.gg/anthropic
• LangChain Discord: https://discord.gg/langchain
• Anthropic Research Blog: https://www.anthropic.com/research
• LangChain Blog: https://blog.langchain.dev/

Navigation

• Previous: Assessment
• Back to Workshop Overview