Most builders use the word 'agent' for everything. There's a critical architectural distinction — and getting it wrong means unpredictable costs, brittle systems, or builds that are far more complex than they need to be.
Most builders use the word 'agent' for everything. There's a critical architectural distinction — and getting it wrong means unpredictable costs, brittle systems, or builds that are far more complex than they need to be.
What's inside
Get instant access
Free. No spam. Unsubscribe anytime.
The word 'agent' gets used for everything — but Anthropic draws a clear architectural line. Workflow agents run through predefined code paths: the sequence is set before execution, cost is predictable, and the same input always produces the same output. Autonomous agents run in a loop: the LLM decides what to do next, how many steps to take, and when to stop. That distinction drives every build decision that follows.
This guide documents the five workflow patterns used in production — prompt chaining, routing, parallelisation (sectioning and voting), and evaluator-optimizer — with real examples for each. It also covers five autonomous agent use cases from Anthropic's research: the SWE-bench coding agent, computer use, open-ended research, outcome-based customer support, and multi-file orchestration.
Includes a cost and reliability comparison table, a warning on runaway autonomous agent spend, and a 4-question decision framework to identify the right type before you write a single line of code.
Who this is for
About the author
Jonathan is a software engineer at Amazon, building Hookem — an AI tool that analyses what makes content go viral — alongside his day job.
Every product decision is documented in public. These guides are the written version of that process.
More guides