Beyond Static Endpoints: Tool Programs as an Interface for Flexible Agentic Web Services

2025-07-28 · Source: cs.SE updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Cloud Computing & IT Infrastructure · Depth: Expert, extended

Summary

ToolPro introduces a novel agent-facing service interface for LLM-based agents interacting with web services, moving "Beyond Static Endpoints." It addresses the limitations of traditional static API endpoints by representing multi-step tool intent as executable tool programs. ToolPro combines constraint-guided program construction, effect-aware replay for "exactly-once" state-modifying calls, and a profile-driven policy to adaptively decide when program execution outperforms stepwise calling. Implemented over MCP-style services with WebAssembly sandboxing, ToolPro was evaluated on real-world applications like Memos, Directus, and MinIO. It demonstrated significant efficiency gains, reducing end-to-end latency by up to 53.4% and client-side traffic by up to 96.1%, with larger improvements under higher network latency and workflow complexity.

Key takeaway

For AI Architects designing agentic web services, consider adopting tool programs to enhance workflow efficiency and reliability. Your systems can significantly reduce end-to-end latency and client-side traffic by consolidating multi-step interactions into single, executable units. Implement effect-aware replay to ensure retry-safe WRITE operations, preventing state corruption during repairs. This approach optimizes resource use and improves agent performance.

Key insights

Tool programs offer an executable, effect-typed interface for LLM agents to manage complex web service workflows efficiently.

Principles

• Represent tool intent as executable programs
• Enforce "exactly-once" semantics for WRITE operations
• Adaptively consolidate interactions based on cost models

Method

ToolPro synthesizes, projects, compiles, and executes programs in a WebAssembly sandbox, using compiler/runtime feedback for bounded in-place repair and logging WRITE outcomes for retry safety.

In practice

• Utilize WebAssembly for secure sandboxed execution
• Implement explicit READ/WRITE effect typing
• Profile network latency to optimize execution mode

Topics

• Agentic Web
• Tool Programs
• LLM Agents
• WebAssembly
• Service Interfaces
• Retry Safety

Code references

• morgen52/toolpro_icml26
• usememos/memos
• directus/directus
• minio/minio
• bytecodealliance/wasmtime

Best for: AI Scientist, Research Scientist, AI Engineer, MLOps Engineer, AI Architect

Related on AIssential

• Beyond Static Endpoints: Tool Programs as an Interface for Flexible Agentic Web Services — ToolPro enables LLM agents to execute complex web service workflows via compact, effect-aware tool programs.
• The Semi-Executable Stack: Agentic Software Engineering and the Expanding Scope of SE — Agentic AI expands software engineering's scope to "semi-executable artifacts" beyond traditional code.
• From Workflows to Multi-Agent Systems: How to Choose — Choosing between workflows, single agents, and multi-agent systems depends on task complexity, dynamism, and cost-benefit analysis.
• AI Didn’t Kill the Web, It Moved in! — Olivier Leplus (AWS) & Yohan Lasorsa (Microsoft) — AI is transforming web development across the entire lifecycle, from coding to agent-centric web app design.
• The End of Software Engineering: How AI Agents Are Fundamentally Restructuring the Software Paradigm — AI agents fundamentally restructure software engineering by making code an ephemeral tool for LLM-driven reasoning, not static logic.
• One Flexible Tool Beats a Hundred Dedicated Ones — Flexible CLI tools outperform dedicated MCP tools for LLM agents due to improved compositionality and reduced context bloat.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.SE updates on arXiv.org.