Agentic AI Harness Pattern

The Top 15 Patterns Learned from Leaked Claude Code and Hermes Agent

By: Ken Huang, Grace Huang

Narrated by: Virtual Voice

Most AI tutorials teach you prompts. This book teaches you patterns.



Production AI engineering — the discipline of turning a language model into something reliable, safe, auditable, and shippable — is mostly undocumented. The libraries churn every quarter. The patterns endure.



Agentic AI Harness Pattern distills 15 of those patterns by reading two mature production codebases side by side: Claude Code, Anthropic's TypeScript CLI for agentic coding, and Hermes, a Python agent built to run across messaging platforms. The two systems make different language choices, different concurrency choices, and different deployment choices — but the harness pattern they implement is the same.



Every chapter follows the same rhythm:

• 
  
• Name the pattern. What problem is the harness solving?
• 
  
• Show Claude Code's implementation in real TypeScript.
• 
  
• Show Hermes's implementation in real Python.
• 
  
• Compare them as a table. Where do they diverge, and why?
• 
  
• Recommend when to use which. A decision rule, not a hot take.
• 
  
• Apply the pattern to a defensive cyber-security agent. A worked example that shows the pattern under operational pressure.
• 
  
  
  

Inside the 15 patterns

1. 
   
2. The Harness Paradigm — why a model alone is not a product
3. 
   
4. Tool Architecture and the Tool Contract — the boundary between reasoning and consequence
5. 
   
6. The Query / Agent Loop — what happens between the model's tool call and the next turn
7. 
   
8. Permission Systems and Safety Guardrails — gating the destructive set
9. 
   
10. Tool Orchestration and Execution — partitioning safe vs. serial work
11. 
    
12. Context Management at Scale — the five strategies before compaction
13. 
    
14. Multi-Agent Coordination — when one agent isn't enough
15. 
    
16. Memory Systems and State Persistence — three tiers, one cache
17. 
    
18. Observability and Debugging — distributed tracing for non-deterministic systems
19. 
    
20. Production Deployment Patterns — SDK-first vs. gateway-first
21. 
    
22. Hook / Event-Driven Automation — the layer above the loop
23. 
    
24. The Skill System Pattern — capabilities as content, not code
25. 
    
26. MCP Integration — connecting agents to the world
27. 
    
28. Model Routing and Provider Abstraction — falling back without falling over
29. 
    
30. Structured Output and Schema-Constrained Generation — when free text isn't enough
31. 
    
    
    

Who this book is for

• 
  
• Engineers building AI products who keep hitting the same architectural questions and want vetted answers.
• 
  
• Architects and tech leads making the build-vs-buy-vs-wrap decision for an agent platform.
• 
  
• Security and compliance reviewers who need to understand how a production agent enforces a destructive-action gate, an audit trail, and an iteration budget.
• 
  
  
  

Each chapter stands alone. Read what you need; read end-to-end and the patterns compound. Either way, you'll close the book with a working mental model of how to design an AI agent that survives contact with production.

About the authors

Ken Huang is CEO of DistributedApps.ai, advising organizations on production-grade agent deployment at the intersection of AI, distributed systems, and security.



Grace Huang is a Product Manager and AI Engineer at PIMCO, where she ships AI features for the world's largest fixed-income asset manager. Her focus is the engineering rigor that makes AI products trustworthy in regulated environments.



The model is intelligence. The harness is the system.

Start here.