Ep6 · Planning Agents From Scratch — Plan-and-Execute vs ReAct (Python)

Every agent we've built has been a REACTOR: it decides one step at a time, looking at the last result before choosing the next move. That's ReAct, and it's flexible — but it never steps back to see the whole job. This episode we build the other pattern: PLAN-AND-EXECUTE. The agent reads the task, writes the whole numbered route up front — before any tool runs — then walks it, step by step. The pattern is three moves: • plan — one model call turns the task into a numbered list of steps, each doable with one tool. You can READ the plan before anything executes. • execute — walk the list, run the right tool for each step, and carry the results forward, so a later step can use the numbers an earlier step found. • answer — one last call folds every result into a final reply. ReAct vs plan-and-execute isn't "which is better" — it's a trade. Planning gives you a visible, auditable route and shines when the steps are mostly knowable; ReAct adapts better when each step depends on what the last one turned up. Real agents often do both. NEW this episode: a tiny TRACER. An agent is a multi-step loop that's invisible from the outside — you only see the final answer. We wrap the run in a ~40-line tracer that records every model call and tool call — tokens, cost, latency — prints the whole run as a tree, and saves it to a SQLite file. It's the per-RUN analog of Season 1's per-CALL cost meter, and every agent from here on runs under it. You can't improve what you can't see. ⏱️ Chapters 0:00 Every agent so far just reacts 0:30 ReAct vs plan-and-execute 1:09 The pattern: plan, execute, answer 1:41 New this episode: a tracer 2:14 The code: adding a planner 2:49 The tracer in code 3:32 Planner, executor, answer 4:43 Run it: the plan, then the run tree 7:07 Recap + Ep7 tease (markdown memory) 🧠 What you'll learn • Plan-and-execute vs ReAct — and when each one wins • Writing the whole route up front, then executing with results carried forward • Why the calculator step can use numbers the look-up steps just found • A tiny from-scratch agent tracer: run tree + tokens + cost, persisted to SQLite Same OpenAI SDK, free Gemini key (no cloud bill) as the whole series. 💻 Code (clone & run, free Gemini key): https://github.com/vahid8/ai-agents-f... ▶ Season 1 — AI Engineering from Scratch (the gateway + cost meter this tracer echoes): https://github.com/vahid8/ai-engineer... 🔗 Free Gemini key: https://aistudio.google.com/apikey #AIagents #Planning #ReAct #Agents #Python #LLM #Gemini #Tracing #Observability