VAP: Verified Action Protocol for AI Agent Software Interaction

Описание и применение

VAP предоставляет протокол для проверки действий AI‑агентов, позволяя им точно знать, завершилось ли действие успешно, даже при потере ответа, и обеспечивая восстановление при неопределенности.

Задача

AI agents struggle to determine whether an action succeeded when the tool’s response is lost or uncertain, leading to blind retries or incorrect success assumptions.

Результат

VAP demonstrates verified actions, lost-response reconciliation, and partial batch compensation in a synthetic demo, matching outcomes of a durable intent log client.

Рабочий процесс

• 1. Define tool actions, expected effects, verification checks, and uncertainty handling in VAP schema.
• 2. Agent prepares an action with scoped permissions via VAP.
• 3. VAP records intent before dispatching the action to the tool.
• 4. Tool executes the action; VAP tracks execution separately from verification.
• 5. VAP reads back state and evaluates defined checks to verify the result.
• 6. If verification fails or is unknown, VAP reconciles by checking existing state and may retry or compensate.
• 7. VAP keeps an event history and evidence of what was checked for audit and debugging.

Доказательства

The answer that started this project was: Its current state, available actions and the outcome of each action—in a consistent, verifiable way.

Источник

If the result cannot be established, VAP reports unknown.

Источник

The demo uses a synthetic issue tracker. It deliberately creates an issue and drops the response.

Источник