agentic-engineering

Eval-first AI development ensures code generation quality by defining and running capability and regression evaluations before and after AI agents implement features. This approach uses human oversight for risk management while AI handles the actual coding tasks.

You break down complex software engineering tasks through task decomposition, splitting large features into smaller, manageable 15-minute units. This process ensures each unit is verifiable and can be independently routed to AI agents for implementation.

Model routing optimizes AI development costs by assigning 15-minute task units to appropriate AI model tiers based on complexity. Simpler tasks route to Haiku, while complex implementations use Sonnet or Opus, ensuring cost-aware resource allocation.

Yes, agentic engineering workflows support developing new software features like user authentication modules. AI agents handle the code generation, while engineers use evals to ensure correctness, safety, and quality control throughout the cycle.

You should avoid relying solely on AI agents for task decomposition when a project lacks clear capability and regression evals. Without these evaluations, human oversight cannot effectively manage quality control and risk management for the generated code.