Single-Agent Patterns

• ✓Build a ReAct agent that reasons before every tool call, survives a failing tool, and stops at a step limit
• ✓Build a Plan-and-Execute agent with a planner, an executor and a replanner, and see a stale plan fail
• ✓Build a reflection loop and compare a vague critic, a rubric critic and a rules-as-code critic on the same emails
• ✓Measure tool selection on a scoreboard and fix look-alike tool descriptions
• ✓Put a code guard in front of an irreversible tool so a bad request cannot reach it
• ✓Choose between the four patterns from your own measurements of cost and correctness

Pick the right loop for an agent before you build it, and know the failure each loop brings - the decision that comes before every agent build.