Playbook/Stage 04

Go Deeper

Level up the craft

Agents & Tools

Letting your AI take actions, not just generate text. Tool use, chains, multi-agent patterns.

Builder’s checkAgents are the most exciting and most over-applied pattern in AI right now. Before you build one: show your current product to five users. Most of what people want automated turns out to be a good prompt and a button, not an autonomous agent loop. The most shipped, most paid-for AI products are almost embarrassingly simple under the hood.

An agent is an LLM that can reason about how to accomplish a goal, plan steps, take actions by calling tools, observe results, and adjust. The key difference from a simple API call: agents operate in a loop.

The ReAct pattern

  1. Think. Reason about the current state and what to do next.
  2. Act. Choose and execute a tool.
  3. Observe. See the result.
  4. Repeat or finish.

When NOT to use agents

  • Single-step tasks. If you can do it in one API call, an agent loop is overhead.
  • No external actions needed. If you just need to answer from documents, use RAG.
  • Deterministic output required. Agents are unpredictable by nature.

Agents break in ways that prompts don't. If you build an agent, you need guardrails: maximum step limits, cost controls, and confirmation gates for any action that changes data or costs money. Log everything. And before you build any of this: have you validated that users actually need this level of automation?

Multi-agent systems

Common patterns: Supervisor (one agent coordinates others), Pipeline (agents process in sequence), Debate (agents argue different perspectives). For most products, a single agent with well-designed tools is more than enough.