Select the smallest mechanism

Decompose the intelligence decision

Do not ask whether the product should use an agent. Ask what each consequential step must decide, which evidence it can access, what effect it may cause, how wrong it may be, and how quickly it must respond. Many useful systems combine rules, retrieval, model judgment, and human review without making the whole workflow agentic.

Start with the smallest sufficient mechanism. Deterministic code is appropriate for stable rules and invariants. Optimization is useful when a defined objective and constraints must be balanced. Classical ML fits repeatable prediction with labeled data. Retrieval supplies context but does not decide. Foundation-model calls handle language judgment; bounded agents handle multi-step uncertainty only when the extra autonomy earns its cost and risk.

Compare on operating constraints

For each candidate, compare evidence quality, explainability, latency, cost, change frequency, failure detectability, recovery, and authority. Include the human route as an explicit mechanism, not a vague fallback. If a human must review every case, model the review time and the conditions that make the recommendation useful.

Record what would invalidate the choice: data drift, policy change, cost growth, new deterministic coverage, unacceptable error, or lack of adoption. A mechanism earns continued use; it is not permanent because it was once selected.

Keep the architecture honest

Document route-specific selection before choosing frameworks. The record should allow another engineer to replace a model with a rule or a human step without rewriting the business contract.