Execution Trees, Not More Logs: A Better Debugging Model for AI Agents

Описание и применение

Использование деревьев выполнения вместо плоских логов позволяет явно видеть иерархию операций, отлавливать ошибки, повторные попытки и параллелизм, ускоряя отладку AI‑агентов.

Задача

Flat logs lack explicit causality, making it hard to understand which operations caused others, detect failures, retries, or concurrency in AI agent runs.

Результат

Using execution trees, developers can quickly identify ownership, fallback usage, retries, and parallelism, leading to faster bug detection and more reliable agent behavior.

Рабочий процесс

• 1. Instrument agent code with AgentInspect wrappers (inspectRun, step, step.tool, step.llm).
• 2. Run the agent locally, generating a trace in .agent-inspect directory.
• 3. View the execution tree with npx agent-inspect view to inspect paths, failures, retries, and concurrency.
• 4. Identify patterns (e.g., retries, fallbacks) that indicate bugs or performance issues.
• 5. Encode deterministic checks using agent-inspect check or TraceContract API to enforce invariants (e.g., max retries).
• 6. Integrate checks into CI to fail on violations.

Доказательства

An execution tree makes the relationship explicit: support-agent ├── plan ├── fetch-inventory (failed: 503) ├── fetch-inventory (success) └── draft-answer

Источник

The tree preserves both facts: the primary path failed; the recovery path completed.

Источник