agentic-engineering

AI agent workflows for complex engineering tasks are automated by routing decomposed unit tasks to models like Haiku, Sonnet, and Opus while humans oversee quality, risk, and cost-tracking guardrails.

The evaluation-first loop in agentic engineering defines capability and regression evaluations, captures failure signatures from a baseline, executes implementation, and compares deltas after re-running evaluations.

Complex code implementation tasks are decomposed using a 15-minute unit rule, requiring each unit to be independently verifiable, present a single primary risk, and define clear completion conditions.

Route Haiku for classification and boilerplate, Sonnet for implementation and refactoring, and Opus for architecture, root cause analysis, and multi-file invariants to optimize engineering workflows.

Reviewing AI generated code should focus on invariants, edge cases, error boundaries, security and authentication, hidden coupling, and rollout risks to ensure quality and mitigate potential issues.

You can track token and retry costs for AI agent engineering sessions by logging the model, tokens, retries, real time, and success or failure status for each task to enforce cost discipline.