AI Agents

An agent is a loop with tools, and almost every production problem with agents comes from that loop: it runs too long, calls the wrong tool, loses its state on restart, or takes an action nobody approved. Model quality is rarely the bottleneck.

These articles cover agent architecture from that angle — tool interface design, when a state machine beats a free-running loop, durable checkpointing so a crashed run can resume, human approval gates on destructive actions, and the specific pitfalls that show up once an agent handles real traffic.

Several posts pair agents with retrieval, since most useful agents are also RAG systems.