让大模型工具调用更安全可控，自动验证并智能路由。
Tool Forge: A Validation-Carrying Toolchain for Governed Agentic Execution
- 将自然语言指令转为带验证的可管理工具包，实现全流程治理。
- 工具上下文减少99.2%，微F1达0.901，25项任务全部生成并通过验证。
- 适合企业级AI代理系统构建，关注安全性与可运维性。
大型语言模型代理正被期望执行实际操作：调用API、操作文件、组装工作流及在企业系统中行动。然而，支撑这些操作的工具层仍常被视为手工集成产物或静态的接口列表。本文提出Tool Forge，一个可携带验证信息的工具链，能将自然语言的能力意图转化为受控、沙箱验证、目录化的工具资产，并通过高效的路由层暴露给代理。Tool Forge将工具视为封装了意图、能力契约、实现代码、依赖策略、测试用例、文档、运行时验证证据、生命周期状态、凭证绑定和路由元数据的完整单元。系统引入路由器，按意图范围暴露工具会话，而非将完整目录模式加载至模型上下文。我们描述了系统架构、验证流程、面向MCP的路由模型、治理控制及开源实现的初步可复现基准。在83个路由器基准案例中，Tool Forge路由器实现0.901的综合微F1，相较直接加载全目录模式，任务流工具上下文降低99.2%。在25个本地工具任务的端到端生成测试中，成功生成25个工具包，对确定性验收检查的微F1达0.940，23个通过实时沙箱验证。这些结果作为初始系统基准，非代表当前最优水平。论文还指出了对抗性路由、更广义API对齐、沙箱隔离及跨系统评估等未解决问题。
原文摘要 · Abstract (English)
Large language model agents are increasingly expected to perform operational work: calling APIs, manipulating files, assembling workflows, and acting inside enterprise systems. Yet the tool layer on which this execution depends is still commonly treated as either a hand-written integration artifact or a static list of schemas exposed to a model. This paper introduces Tool Forge, a validation-carrying toolchain for converting natural-language capability intent into governed, sandbox-verified, cataloged tool artifacts and exposing those artifacts to agents through a token-efficient routing layer. Tool Forge treats a tool as a capsule containing intent, capability contract, implementation, dependency policy, tests, documentation, runtime validation evidence, lifecycle state, credential bindings, and routing metadata. It also introduces a Router that exposes intent-scoped tool sessions instead of loading full catalog schemas into the model context. We describe the system architecture, validation pipeline, MCP-facing routing model, governance controls, and initial reproducible benchmarks from the open-source implementation. Across 83 Router benchmark cases, Tool Forge Router achieves aggregate micro-F1 of 0.901 while reducing estimated task-flow tool context by 99.2% relative to naive full-catalog schema exposure. In a 25-case end-to-end generation probe over local-tool tasks, Tool Forge generates 25 of 25 tool bundles, reaches micro-F1 of 0.940 against deterministic acceptance checks, and passes 23 of 25 live sandbox validations. These results are presented as an initial systems benchmark, not as a state-of-the-art claim. The paper identifies remaining challenges in adversarial routing, broader API grounding, sandbox isolation, and cross-system evaluation.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。