ai-agents-architectlisted

# AI Agents Architect Agent architecture is the set of decisions that turn a capable language model into a system that acts reliably: how it perceives, plans, uses tools, remembers, and recovers. This skill covers those decisions as engineering trade-offs. ## Overview Every agent is a loop: perceive state, reason about it, take an action, observe the result, repeat. Around that loop sit the supporting systems — tools with clean interfaces, a planner that decomposes goals, memory that persists context, and guardrails that bound behavior. The architect's job is to pick the right loop topology (single agent, supervisor, swarm), the right planning strategy (reactive vs. deliberative), and the right failure semantics for the task's risk level. ## When to use - Designing a new agent system: choosing loop structure, tools, and memory before writing code. - A prototype agent is unreliable: diagnosing whether the fix is prompting, tool design, planning, or evaluation. - Scaling from demo to production: adding guardrails, observability, cost controls, and human-in-the-loop. - Deciding between single-agent, hierarchical, and multi-agent designs for a complex workflow. ## Core concepts - **Perception-action loop**: the core cycle — state in, reasoning, action out, observation back. Keep the loop observable: log every turn with its inputs and outputs. - **Tool design**: tools are the agent's hands. Each tool needs a clear name, typed inputs/outputs, idempotency where poss