Choose the Simplest System That Passes Eval

Once "good" is defined, there is a temptation to build an agent right away. But a model's strength doesn't cancel architecture - it lets you simplify it. The rule is simple: take the best available model and the minimal harness, and add complexity only in response to a specific, observed failure mode.

It is convenient to navigate by the signal of the task - it tells you where to start and what not to add yet.

• One input and one verifiable result - a single call with a schema is enough. An agent loop is redundant here.
• You need domain knowledge - retrieval plus one call, not a research team of agents.
• There are clear sequential stages - a workflow in code, not an LLM orchestrator for the sake of step order.
• You need to pick one of a few stable routes - a classifier or router, not free planning.
• The path changes with intermediate discoveries - an agent with tools and max steps, but not unlimited autonomy.
• There are independent expensive branches - parallel workers and a synthesizer, not a shared chat of all agents.

The logic in every line is the same: add exactly the mechanism the task demands, and not one more.

The strong model first. Anthropic's heuristic: run a minimal prompt on the best model, then add only what fixes the errors you found. Once the system has stabilized, you can measure which requests are safe to hand to a cheaper model.