Define and measure
Name the outcome, inputs, stakeholder, failure impact, representative examples, and success metrics before architecture.
Do you actually need an AI agent?
Start with the problem, not the architecture. Answer a few questions to find the lowest-complexity system that can do the work reliably.
Explore the complete decision tree and design guideDecision 1 of up to 6 17%
Start with the simplest system
Can explicit rules solve the task reliably?
For the same input, could code follow the same rules and produce the correct output every time?
After the decision
Design the complete system
The recommendation chooses an implementation family. Reliability comes from the design loop around it.
Decompose and classify
Give each part an input, output, trigger, required context, failure mode, and verification method.
Compose with contracts
Connect parts as a chain, DAG, parallel fan-out, or runtime loop. Define schemas, retries, timeouts, and ownership.
Evaluate and improve
Measure quality, cost, latency, safety, and recovery. Improve the weakest step at the lowest useful complexity.
A useful routing rule
Let the model classify. Let code route.
If you already know the available categories and what each one should trigger, return a constrained label or schema from the model and handle the next step in code.
This keeps agent descriptions out of the prompt, reduces ambiguity and token cost, and makes routing easier to test. Reserve an orchestrator for the genuinely unknown or exceptional path.