Resources
Guides
Design questions that arrive the moment a language model stops recommending and starts doing. Each guide opens with the answer, says who it is for, works through an example, is explicit about what the argument rests on, and ends with what it does not cover.
Designing agentic systems
The six decisions that make up an agentic system's design, in the order they are worth making, worked through one example.
Read the pillarWhen a workflow is enough
An agent is a decision about control flow, not about capability. The test is whether you can write the steps down before the request arrives.
Answers When should an engineering team use an agent instead of a fixed workflow?
Read thisWho may call the tool
Permissions live at the tool, not in the prompt. One job per tool, limits it enforces itself, and its own check on the evidence it was handed.
Answers How should an agent's tool permissions be designed?
Read thisWhat to do with uncertain evidence
Three outcomes rather than two — act, refuse, and hand over with the gap named. Uncertainty is a state to route, not a number to threshold.
Answers What should an agent do when its evidence is uncertain?
Read this4 guides are published. Further clusters — evaluating agent behaviour, reliability and cost, and running a design review — are planned and are not written yet. Nothing appears on this page until it has been written and is ready to be read.