agentic-engineering

An eval-first loop baselines AI agent capability and runs regression checks before implementation. This approach ensures engineering tasks yield verifiable results through unit-level checks, preventing regressions when scaling from small features to multi-file systems.

Decompose engineering tasks using the 15-minute unit rule to create independently verifiable units. Each unit should have a single dominant risk, allowing AI agents to implement work reliably and ensuring outcomes are checked through unit-level validation.

Model routing assigns engineering work to Haiku, Sonnet, and Opus tiers based on task complexity. This cost-aware strategy maximizes quality and speed by matching the AI agent's model tier to the specific requirements of the decomposed unit.

Track costs by monitoring per-task metrics including model tier, token estimates, retries, and wall-clock time. This cost discipline ensures governance and transparency when scaling AI agents across multi-file engineering systems.

Yes, this approach scales from small features to multi-file systems by applying risk-based review checkpoints. It ensures verifiable results through regression evaluation while maintaining cost discipline across different levels of engineering complexity.

Cost-aware model routing requires balancing quality and speed across Haiku, Sonnet, and Opus tiers. Limitations include managing risk-based review checkpoints and ensuring regression evaluation catches errors when routing complex multi-file engineering tasks.