Lesson 4 of 6

Human-in-the-loop (Crawl-Walk-Run)

7 min read

The instinct with a capable model is to build one autonomous agent and let it rip. The instinct is usually wrong — and expensive to unlearn in production.

Crawl, walk, run

Autonomy is a dial you turn up slowly, not a switch you flip. Crawl: a fixed, predictable path where you decide every step. Walk: the flow branches and parallelizes, but still on rails you designed. Run: the system loops and self-corrects. Most valuable AI features live at crawl and walk — a [workflow](glossary://workflow) you can trust, not a free-roaming agent.

A fixed workflow keeps a human in the loop by design: you own the path, so nothing runs somewhere you didn't route it.

Pick the smallest shape that works

Each pattern is a shape you can reason about: a straight chain, a router that sorts the request, parallel branches that merge, an orchestrator that delegates, an evaluator loop that checks its own work. Match the smallest shape to the job. A predictable path is easier to test, to debug, and — crucially — for a person to supervise than an open-ended agent.

Reach for the simplest workflow shape that does the job. Predictable beats clever when a human has to stay in control.

'Run' isn't the goal line — it's the exception. Turn autonomy up only when the task genuinely needs it and the cost of a wrong step is low. Most of the time, crawl and walk ship better products.