3 · Tools, Agents & MCP › Phase 4 › Lesson 1 of 6

Planning strategies that survive step 30

What this lesson covers

An agent that plans, then acts, then re-plans behaves qualitatively differently from one that reacts step by step. Neither is the right answer everywhere.

The outline

1. ReAct. Reason-then-act, one step at a time. The default. When it works and when it doesn’t.
2. Plan-and-execute. Draft a plan up front, then execute. Why this is often better for structured tasks, and its characteristic failure mode.
3. Reflection. Ask the model to critique its own plan or output before continuing. What it buys you, what it costs, and when it’s theatre.
4. Depth-first vs. breadth-first. The tradeoff most agent frameworks don’t let you tune.
5. State a plan can revise. Why an immutable plan is a bug and how to let the model change its mind.
6. The 30-step problem. Why every planning strategy degrades past a certain depth, and what to do about it.

Coming soon

In outline.