agentic-engineering

Route AI tasks by assigning Haiku for classification or boilerplate transforms, Sonnet for implementation or refactors, and Opus for architecture or root-cause analysis. This model routing strategy balances capability against cost for each agent-sized unit of work.

An eval-first loop defines capability and regression evals, runs a baseline, executes implementation, and re-evaluates deltas. This process ensures AI-driven engineering workflows maintain quality through iterative evaluation from design to implementation.

Break engineering tasks into agent-sized units with clear completion criteria. Task decomposition allows AI agents to perform implementation work while humans enforce quality controls and monitor risk throughout the staged workflow.

Track token usage, retries, and wall-clock time across model tiers to maintain cost discipline. Monitoring these metrics alongside model routing decisions optimizes efficiency and exposes operational costs for AI-driven engineering workflows.

Yes, staged decomposition scales refactoring by assigning Sonnet for implementation and Opus for root-cause analysis. The eval-first loop regression tests ensure that large-scale changes maintain capability baselines throughout the iterative process.

Eval-first execution prevents regression by establishing baselines before implementation. Direct AI implementation risks quality degradation, whereas iterative evaluation enforces human quality and risk controls over agent-generated code.