agentic-engineering

Evaluation-first planning applies capability evaluation and baseline comparison before execution to ensure correctness. This mechanism defines unit-level finish criteria and regression checks, allowing iterative improvement and rollback safety in agent engineering.

Task decomposition breaks complex requests into agent-friendly steps based on complexity, distributing them to appropriate model layers. This hierarchical routing enhances scalability and maintains quality control through session management and phase switching.

Risk control and quality gates are managed through session strategies that include phase switching, session resumption, and regression checks. By defining unit-level finish criteria, the workflow enforces quality control and enables repeatable execution across stages.

Yes, hierarchical model routing distributes decomposed planning steps to models like Haiku, Sonnet, and Opus based on task complexity. This structured routing improves execution stability and ensures appropriate model allocation for each agent unit.

An evaluation-first workflow is less suited for simple, single-step tasks that do not require cross-model planning or baseline comparisons. Without the need for regression checks, hierarchical routing, or iterative improvement, this structured risk management adds unnecessary overhead.