I asked what would make software easier for AI agents to use. Then I built VAP (Feedback Wanted).

The article discusses the need for software to provide a consistent, verifiable way for AI agents to understand the actions they can perform and the outcomes of those actions. The author proposes the Verified Action Protocol (VAP) as a solution to this problem.

Key points:

- Most discussions about improving AI agents focus on the model itself, but the author believes that improving the software agents interact with could be equally beneficial.

- VAP is an open-source protocol that defines how tools should expose their actions, expected effects, verification methods, and how to handle uncertain outcomes.

- It sits between an agent and its registered tools, allowing developers to describe actions, their effects, verification checks, and how to investigate uncertain results.

- VAP provides a shared place to prepare actions, record intents, track execution and verification, read back state, reconcile uncertain operations, and keep an event history.

- The demo uses a synthetic issue tracker to demonstrate VAP's functionality, showing how it handles dropped responses, partially applied batches, and compensation as separate operations.

- VAP does not promise exactly-once execution, but it does provide a structured way for agents to understand actions and their effects, including uncertainty.

- The author is seeking feedback on whether VAP adds value to existing agent workflows, where it would fit or fail to fit, and which real tools developers would want to connect first.