Research map / Multi-agent systems and games / LLM agents
Beyond Static Endpoints: Tool Programs as an Interface for Flexible Agentic Web Services
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ToolPro, which has agents submit an executable, effect-typed tool program (run in a Wasm sandbox) instead of stepwise endpoint calls, with effect-aware replay and a profile-driven policy choosing program vs stepwise mode.
Problem
LLM agents invoking web services via static endpoints must externalize long-horizon workflows with loops, conditionals, and retries as many round trips, inflating latency and client traffic and risking duplicated writes on retry.
Approach
Represent an agent's tool intent as a single executable tool program with explicit READ/WRITE effect annotations, compiled and repaired server-side in a Wasm sandbox; WRITE calls are replayed from a per-intent log to give exactly-once semantics across repair-driven re-executions, and a cost model decides when program mode beats stepwise calling.
Key results
- ToolPro reduces end-to-end latency by up to 53.4% (cbench4: 52.68s to 24.54s) and client-side traffic by up to 96.1%.
- On cbench1-cbench3, ToolPro cuts average latency from 30.16s to 17.91s (40.6%) and raises task accuracy from 0.60 to 0.93.
- Disabling replay raises average latency from 17.92s to 21.45s and fallback rate from 0/15 to 3/15.
Surprising findings
- Go-based program generation reached only 0.40 accuracy on cbench2 versus 0.80 for Rust, indicating compilation brittleness across languages.
- At low RTT (1-100ms), stepwise mode beats program mode, so consolidation is not always beneficial.
Limitations
- Program-mode failure rate depends on LLM coding capability (qwen3-coder-flash 80% vs 100% for gpt-5.1 and gemini-3-flash-preview on cbench2).
- Replay requires idempotent or deterministic endpoints; nondeterministic non-idempotent endpoints force fallback to stepwise calling.
- Evaluation uses only three open-source applications and small workflow sizes (N up to 20) with five runs averaged.
Experimental setup
Datasets: Memos Directus MinIO cbench1-cbench4 (constructed supplemental workflows)
Benchmarks: Memos.r Memos.w Directus.r Directus.w MinIO.r MinIO.w cbench1-cbench4
Metrics: end-to-end latency client-side traffic task accuracy
Baselines: Stepwise MCP Web Service (MWS) ToolPro-step ToolPro-prog
Builds on: MCP CodeAct GraphQL ORFA
Closest papers by meaning
- General Agent Evaluation (NeurIPS 2026 · similarity 0.85)
- ComplexMCP: Evaluation of LLM Agents in Dynamic, Interdependent, and Large-Scale Tool Sandbox (ICML 2026 · similarity 0.85)
- OSWorld-MCP: Benchmarking MCP Tool Invocation In Computer-Use Agents (ICLR 2026 · similarity 0.85)
- MCPMark: A Benchmark for Stress-Testing Realistic and Comprehensive MCP Use (ICLR 2026 · similarity 0.85)
- TimeWarp: Evaluating Web Agents by Revisiting the Past (NeurIPS 2026 · similarity 0.84)
- ContractBench: Can LLM Agents Preserve Observation Contracts? (NeurIPS 2026 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.