Agent Harnesses and the Architecture of Software

Описание и применение

Пример демонстрирует, как агентский harness координирует вызовы инструментов REA для анализа JavaScript‑приложения, сохраняя доказательства и контекст между шагами.

Задача

How can an AI agent orchestrate tool calls to perform reverse‑engineering tasks while preserving evidence and context across multiple steps?

Результат

The example demonstrates that an agent harness can coordinate tool calls to a reverse‑engineering toolkit, preserving evidence across steps and enabling iterative investigation of application features.

Рабочий процесс

• 1. Model receives user request to investigate search functionality.
• 2. Agent harness proposes tool call analyze_javascript_application to REA.
• 3. REA performs static analysis and returns a summary with a reference to retained evidence.
• 4. Model requests a paged view of modules via inspect_analysis_view using the evidence ID.
• 5. Model identifies module search.js and calls trace_application_feature with the same evidence ID and module seed.
• 6. REA returns a feature trace, possibly revealing HTTP requests or local indices.
• 7. Model evaluates trace, may request dynamic execution evidence if needed.
• 8. Agent harness logs all messages, tool calls, and results for persistence and recovery.

Доказательства

The model receives the request and the available tool descriptions. It might propose analyze_javascript_application. The surrounding agent runtime dispatches that call to REA.

Источник

The model can then ask to see a page of modules: …

Источник

The trace might lead toward a local index, or toward code that sends an HTTP request. Those clues would lead to different follow‑up work.

Источник