Production AI agent architecture: Anthropic’s complete 13-block system

A production AI agent is not just a set of instructions in one long prompt. It's a full-featured system where each component handles a separate aspect of operation: resource management, OODA loop, responsibility distribution between agents, error handling, monitoring. Without clear architecture, your agent will behave unpredictably.

What

Anthropic Revealed The company published two real production prompts used in live systems: research_lead_agent.md and research_subagent.md. These are not textbook examples, demonstrations, or simplified versions. These are real prompts that handle actual work. When you break them down, you see an architecture that isn't taught in prompt engineering courses. The key discovery: instead of one magical long prompt, Anthropic uses a system of 13 clearly defined, independent blocks. Each block handles one task and can be reused in other agents. This means the architecture scales not through rewriting, but through composition.

13

Blocks Every Production Agent Needs Here's the complete checklist of components that must be present in any production agent: Role and persona — defining how the agent thinks and behaves Goal and context — clear specification of what needs to be done and why Call budget and resource ceiling — the threshold for when to stop OODA loop — the Observe→Orient→Decide→Act cycle that repeats Error handling and fallback strategies — actions when something fails Responsibility distribution between agents — dividing the work Tools and their limitations — what the agent can and cannot use Output format and result validation — how to verify correctness Monitoring and logging — a way to see what happens inside Timeouts and retry logic — preventing hangs and repeated failures Self-checking and verification — the agent verifies itself before output State normalization — formatting the result into readable form * Guardrails and safety layer — preventing boundary violations This is not optional and not "nice to have". If even one component is missing, the agent will behave unstably and unpredictably. I've seen this dozens of times — agents lose context, exceed budget, hang on edge cases.

Lead and Sub: Division of Labor The difference between a lead agent and a sub agent is fundamental.

Lead is a strategist and coordinator. It sees the entire big picture, plans strategy, makes key decisions, manages the work of a whole team of sub agents. Sub agent is a narrow specialist. Fast, cheap to call, takes one specific, well-defined task and reports the result. Sub listens to lead and shows no initiative. Anthropic uses this scheme because it scales. One lead can coordinate five sub agents. Then you add five more — the lead handles them. Five more — everything works. You don't rewrite the lead, don't change the architecture. Just expand the team.

Ready

Template in Hand After analyzing Anthropic's prompts, I extracted patterns and created a SKILL.md template. Using this template, you can go through your agent and make sure you haven't missed anything. It's like a pilot's checklist before takeoff, but for AI agents. The key rule: use it BEFORE going to production. Because adding components "on the fly" and rewriting prompts in live conditions is a nightmare. Spend one day on architecture — then you'll be grateful for months.

What

This Means If you're building an agent right now, this architecture can save you weeks of debugging. Because usually all components appear in code randomly when disaster has already struck — the agent hung, exceeded budget, got stuck on an edge case. Now you have a tested template. Use it.