An agent layer is one of five dimensions for analyzing, designing, and building an AI agent. Together, the layers turn an agent from a prompt-driven demo into a system with defined responsibilities, capabilities, reasoning, memory, and feedback. section 1.5 section 11.5
Treat the agent as five stacked layers, then optimize each layer independently before integrating them:
- Persona: defines the agent’s role, boundaries, scope, and explicit “I don’t know” behavior. section 1.5 section 11.5
- Tools and actions: provides focused, typed capabilities and controls when those capabilities run. section 1.5 section 11.5
- Reasoning and planning: structures nontrivial work through plans, execution steps, iteration limits, and self-review. section 1.5 section 11.5
- Knowledge and memory: retrieves relevant information while separating short-term session context from long-term stored memory. section 1.5 section 11.5
- Evaluation and feedback: traces behavior, automates evaluations, collects user feedback, and reviews complex cases. section 1.5 section 11.5
A layer-by-layer design pass
Start with a persona written as an API contract. Add a single-responsibility tool with typed parameters and a structured output. Give the agent a short planning checklist and an iteration cap. Retrieve knowledge through a retrieval tool while pruning session and long-term memory separately. Finish by tracing prompts, tool calls, latency, token counts, and outcomes, then run curated evaluations after changes. section 11.5
