Begin with the workload
Start with a representative task and define a correct result. Identify the inputs, data boundaries, tools, and decisions involved. This makes it possible to distinguish a model-quality problem from a retrieval, coordination, or execution problem.
Design context deliberately
Context is a limited working space. Select evidence for the task, retain its provenance, and separate persistent knowledge from temporary execution state. More context is useful only when it helps the model make a better decision.
A harness makes context, tools, execution, and review part of one observable system.
Coordinate execution
Tool calls, queues, and agent handoffs need explicit ownership and state transitions. Independent work can run in parallel; dependent actions must preserve ordering. Retries should not repeat irreversible side effects.
Evaluate the whole path
Assess response quality alongside token use and end-to-end latency. Keep a representative evaluation set and inspect execution traces when results change. The goal is a system whose behavior can be understood and improved.