arXiv:2606.19992: Beyond Static Endpoints — ToolPro introduces executable programs as an interface for agentic web services

Why static API endpoints are not enough for AI agents

Static API endpoints — predefined calls that return a fixed response — work well for simple queries but break down when an LLM agent needs to execute a more complex flow: a polling loop waiting for a result, conditional branching, or repeating a call after an error. Researchers Mugeng Liu, Shuoqi Li, Yixuan Zhang, and Yun Ma in paper arXiv:2606.19992 identified exactly this gap as the fundamental problem in scaling agents to real-world web services.

What is ToolPro and how does it solve the problem?

ToolPro (Tool Programs) is a system that, instead of static endpoints, offers agents executable programs — compact multi-step descriptions of web service interactions that encode loops, conditionals, join operations, and retry logic. Three key components: constraint-guided construction (building programs under semantic constraints), effect-aware replay (a mechanism ensuring every command with side effects executes exactly once, not more), and an optimal execution-scheduling policy. The implementation uses MCP-compatible (Model Context Protocol) services and a WebAssembly sandbox for isolated, secure program execution on the service side — not the client side.

Results and comparison

Compared to the classic static-endpoint approach, ToolPro achieves up to 53.4% lower end-to-end latency and up to 96.1% less client-side network traffic. The gains are most pronounced under poor network conditions and complex workflows — precisely the scenarios that are most demanding for AI agents. The paper was submitted on June 18, 2026 as part of the Monday arXiv batch.