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

Описание и применение

Деревья выполнения позволяют разработчикам видеть структуру и причинно-следственные связи в работе AI‑агентов, выявлять ошибки, повторные попытки и параллельные операции, а также создавать детерминированные проверки.

Задача

Flat logs obscure causal relationships in AI agent runs, making it hard to identify ownership, failures, retries, and concurrency.

Результат

Execution trees provide a concise, causal view of agent runs, revealing ownership, failures, retries, and concurrency, and enabling deterministic checks that improve debugging and reliability.

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 <run> --dir .agent-inspect --summary.
• 4. Identify suspicious shapes (nested failures, fallbacks, retries, parallel siblings).
• 5. Encode deterministic checks using agent-inspect check or TraceContract API.
• 6. Run checks in CI to enforce invariants.

Доказательства

Execution trees make the relationship explicit: support-agent ├── plan ├── fetch-inventory (failed: 503) ├── fetch-inventory (success) └── draft-answer

Источник

npx agent-inspect view travel-planner --dir .agent-inspect --summary

Источник