Autonomous Agents

This covers the architecture and reliability patterns for building AI agents that can act independently: ReAct loops, plan-execute separation, goal decomposition, and reflection patterns. The emphasis is ruthlessly practical. It pushes you to start with constrained, single-purpose agents and earn autonomy through proven reliability rather than building general-purpose systems that fail unpredictably. The sharp edges table is particularly useful, calling out the compounding error problem directly: 95% success per step becomes 60% by step ten. You'd reach for this when you need to move beyond simple LLM calls into multi-step autonomous behavior, but want to avoid the common trap of demos that impress but never make it to production.