Which Agent Design Pattern Should You Use?

Six shapes can carry the same task. They cost differently, they fail differently, and picking the wrong one is expensive in a way you won't notice for months. This walks through all six and gives you one question that sorts them before you write any code: can you count the steps before you run it? If you can, you have a pipeline you can draw, price and debug. If you can't, the model is deciding how long to work, and that one property drives your cost, your failure modes and your whole debugging story from there on. Built for engineers preparing for AI and agent system-design interviews, and for anyone about to build their first production agent. It ends on a decision procedure you can screenshot. *What's covered* The single-agent baseline, and why skipping it means you never learn whether your workflow helped Sequential, and the stage that fails quietly while the last stage takes the blame Parallel, sectioning versus voting, and why the merge is where these most often go wrong Routing, cascading classifiers, and what better examples did for accuracy in Anthropic's own worked example The pivot: what changes the moment execution stops being predetermined Evaluator-optimizer, its two preconditions, and the stopping criteria you set before you start Why you separate the generator from the evaluator, and what a generous evaluator does to the loop Orchestrator-workers versus parallelization, which is the interview question this answers The autonomous loop, and why observe is the step that carries it, and what you bound it with Nesting, and the rule that your bill inherits the loosest bound anywhere inside it Sources and dating The technical examples and source material in this video come from Anthropic's engineering blog, their product blog, and their official platform documentation, and each source is dated on screen. **The framework connecting those patterns is my own synthesis**, and that includes the count-the-steps question, the cost-predictability framing, and the five-question decision procedure. This is a best-effort snapshot of the latest published material as of 8 September 2026. Anthropic's guidance and APIs change, so check the documentation and the cookbook before relying on any specific figure or API detail. Common workflow patterns for AI agents, and when to use them (Claude by Anthropic, 5 Mar 2026) https://claude.com/blog/common-workfl... Ticket routing use-case guide (Claude Platform Docs, living document, read 8 Sep 2026) https://platform.claude.com/docs/en/a... Building multi-agent systems: when and how to use them (Claude by Anthropic, 23 Jan 2026) https://claude.com/blog/building-mult... Effective harnesses for long-running agents (Anthropic Engineering, 26 Nov 2025) https://www.anthropic.com/engineering... Two things are worth flagging rather than leaving implicit. **The 71 to 93 percent figure is Anthropic's own worked example**, from the ticket routing guide, and it is one result on one set of tickets rather than a number you should expect on yours. The guide is a living document with no publication date, so the date given is the date I read it. **The analysis in this video is mine, not theirs, wherever it goes past what those four documents say.** That includes the count of six patterns, the three consequences of a run-time execution path, and the rule about bills inheriting the loosest bound. The sources are named on screen for the claims they carry. 00:00 Six pictures on the whiteboard 02:06 Try it as one agent first, and keep the score 04:16 Sequential, and the stage that fails quietly 06:56 Parallel, and why the merge is the hard part 10:30 Routing, cascading classifiers, and better examples 15:12 The pivot: when the path is decided at run time 20:44 Evaluator-optimizer, and knowing when to stop 24:07 Separating the generator from the evaluator 26:43 Orchestrator-workers vs parallelization 28:58 The autonomous loop, and the observe step 32:56 Nesting, and the loosest bound 34:52 The decision card 37:22 Back to the ticket 38:52 Count them, or bound them