agent-harness-construction

To improve agent completion rate, stabilize tool naming, constrain input schemas, and enforce deterministic output shapes. Apply observation contracts with status, summaries, next actions, and artifacts to ensure structured and reliable tool-using agent workflows.

An observation and recovery contract enforces tool responses to include clear status, summaries, next actions, and artifacts. It provides explicit root-cause hints, safe retry instructions, and stop conditions to ensure robust handling of failures across exploratory and deterministic phases.

Structure deterministic tool inputs and outputs by constraining input schemas and returning deterministic output shapes. Avoid overlapping or catch-all tools to stabilize the action space, ensuring the agent receives consistent data formats for reliable processing.

Choose between ReAct and function calling by evaluating convergence needs and context budgeting. Apply practical rules for keeping system prompts minimal, loading guidance on demand, and selecting hybrid patterns to optimize architecture selection for better agent completion performance.

Benchmark agent task success by tracking completion rate, retries, pass@1, pass@3, and cost per success. Flag common failure modes and anti-patterns during exploratory and deterministic phases to measure and optimize error recovery behavior effectively.

Agents get stuck in retry loops when lacking explicit stop conditions and safe retry instructions. Implement an error recovery contract with root-cause hints and clear next actions within observation formatting to prevent infinite loops and ensure robust failure handling.