𝗩𝗲𝗿𝗶𝗳𝗶𝗮𝗯𝗹𝗲 𝗔𝗜 𝗔𝗴𝗲𝗻𝘁 𝗘𝘅𝗲𝗰𝘂𝘁𝗶𝗼𝗻 AI agents are becoming capable of handling tasks such as analysing market data, interacting with DeFi protocols, and making decisions based on information from external sources. 𝘉𝘶𝘵 𝘩𝘰𝘸 𝘤𝘢𝘯 𝘸𝘦 𝘬𝘯𝘰𝘸 𝘸𝘩𝘦𝘵𝘩𝘦𝘳 𝘢𝘯 𝘢𝘨𝘦𝘯𝘵 𝘢𝘤𝘵𝘶𝘢𝘭𝘭𝘺 𝘥𝘪𝘥 𝘸𝘩𝘢𝘵 𝘪𝘵 𝘸𝘢𝘴 𝘴𝘶𝘱𝘱𝘰𝘴𝘦𝘥 𝘵𝘰 𝘥𝘰? An agent might report that it checked a price before placing a trade. however, seeing the final result does not prove that the price was checked or that the required steps were followed. This creates a problem for applications that rely on AI agents to perform important tasks. they need a way to check not only the final result but also the evidence behind the actions that led to it. Many of these actions involve API calls. For example, an agent might: ⇒ Send a request to an AI model to generate a decision. ⇒ Query a price oracle to obtain market information. ⇒ Retrieve data from an external service before taking action. If these steps are required by the application’s rules, there needs to be a way to verify that they actually happened. Cryptographic proofs can help verify specific API interactions. 𝘞𝘩𝘢𝘵 𝘮𝘢𝘬𝘦𝘴 𝘢𝘯 𝘢𝘨𝘦𝘯𝘵’𝘴 𝘢𝘤𝘵𝘪𝘰𝘯𝘴 𝘷𝘦𝘳𝘪𝘧𝘪𝘢𝘣𝘭𝘦? @primus_labs uses zkTLS to help verify data obtained from web services. a prover can generate cryptographic evidence of data exchanged with a server over a TLS-secured connection, without revealing all the underlying session data or private credentials. For an AI agent, this can provide evidence of important interactions during its execution. The process can be organised into a few steps: ⇒ 1. Set the rules: Define which actions the agent must perform, such as checking a price before placing a trade. ⇒ 2. Record the required interactions: Generate proofs for relevant API requests and responses as the agent carries out its task. ⇒ 3. Verify the proofs: A verification system checks whether the evidence meets the required conditions. ⇒ 4. Check the execution policy: The application checks whether the required steps were completed and whether their order and results meet its rules. The final step matters. proving an API request occurred doesn’t automatically mean the agent followed its execution policy. The application must still check the evidence against its rules. 𝗪𝗵𝘆 𝗱𝗼𝗲𝘀 𝘁𝗵𝗶𝘀 𝗺𝗮𝘁𝘁𝗲𝗿? As AI agents handle more tasks in DeFi and other automated systems, verifying their actions becomes increasingly useful. Applications can use verifiable evidence of specific API interactions to check whether required steps were completed. This can support auditing and make parts of an agent’s activity easier to verify. the level of verification depends on how the system is implemented.