Agent Tool Protocol

Agent Tool Protocol (ATP) is a protocol designed to enable AI agents to write and execute code directly, rather than relying on predefined tool calls. It addresses challenges encountered in traditional server protocols like MCP, which require developers to build gateways that aggregate multiple services and manage complex stateful connections. com's MCP gateway, drawing on experience with large-scale API integrations and developer infrastructure.

A central feature of ATP is its approach to code execution for AI agents. Instead of loading all available tools into an agent's context before reasoning begins, ATP allows agents to dynamically search for and use only the relevant APIs as needed. This reduces unnecessary context bloat and improves efficiency. The protocol leverages the OpenAPI standard for REST APIs, enabling agents to interact with well-defined endpoints and schemas, and to compose operations such as filtering and mapping data through code execution.

ATP incorporates several layers of security and control to address the risks associated with code execution. Code runs in an isolated environment with no access to the file system, network, or environment variables, mitigating the risk of data exfiltration or unauthorized access. APIs can be annotated as destructive, sensitive, or safe, and there are explicit controls to allow or block specific endpoints. For operations that could alter or delete data, ATP supports human approval workflows, ensuring that critical actions require explicit authorization.

The protocol also supports advanced agent architectures, including multi-level agents that can spawn sub-agents for parallel processing of tasks. This flexibility is intended for developers building scalable and secure AI agent systems that need to aggregate and manage multiple APIs efficiently. ATP is positioned within the frameworks and SDKs class, focusing on enabling dynamic, code-based API composition for AI agents.