Research map / Multi-agent systems and games / LLM agents
I-PTC: Interactive Programmatic Tool Calling for Stateful Tool-Augmented Agents
Huanzhi Mao, Chengkun Cao, Shuo Yuan, Joseph Gonzalez
Read the paperCheck your idea against 137,217 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 Interactive Programmatic Tool Calling (I-PTC), extending programmatic tool calling to stateful execution so agents can reuse intermediate results across tool-use programs.
Problem
Existing programmatic tool calling (PTC) approaches are largely stateless, which limits reuse of intermediate results and interaction across steps in tool-augmented agents.
Approach
Extends programmatic tool calling with interactive, stateful execution so that code-expressed tool use (loops, batching, abstraction) can persist and reuse intermediate results; the abstract does not give further details.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use programmatic tool calling instead of traditional tool calling | GPT 5.4 · PTC-Bench | pass rate: 18.6% -> 71.5% | Improved |
| make PTC stateful and interactive (I-PTC) | GPT 5.4 · PTC-Bench | pass rate: 71.5% -> 80.0% | Improved |
| make PTC stateful and interactive (I-PTC) | Claude Opus 4.7 · PTC-Bench | pass rate: 84.2% -> 82.9% | Hurt |
| make PTC stateful and interactive (I-PTC) | Qwen 3.5-397B-A17B · PTC-Bench | pass rate: 75.3% -> 71.8% | Hurt |
| use I-PTC instead of PTC | Claude Opus 4.7 · PTC-Bench | raw tokens: 184K -> 85K | Improved |
| add persistent cells to stateless execution (ablation) | GPT 5.4 mini and Gemini 3.1 Flash-Lite · PTC-Bench | pass rate: +0.6 to +5.7 points | Improved |
| add on-demand tool inspection (ablation) | I-PTC on GPT 5.4 mini · PTC-Bench | pass rate / raw tokens: 61.3 -> 68.1; 129K -> 44K | Improved |
| add self-evolving API documentation (ablation) | I-PTC on GPT 5.4 mini · PTC-Bench | pass rate: 61.3 -> 62.9 | No clear effect |
| add documentation to stateless PTC (ablation) | PTC on GPT 5.4 mini · PTC-Bench | pass rate: 59.1 -> 57.2 | Hurt |
| delay documentation updates (larger batch size) (ablation) | self-evolving docs · PTC-Bench main | API errors | Hurt |
| use I-PTC instead of PTC on multi-turn | GPT 5.4 mini · PTC-Bench multi-turn | pass rate: 69.0% -> 77.1% | Improved |
| use I-PTC instead of PTC on error recovery | GPT 5.4 mini · PTC-Bench error-recovery | pass rate: 61.3% -> 72.9% | Improved |
Experimental setup
Builds on: programmatic tool calling
Closest papers by meaning
- Toward Interactive Understanding of Code APIs (NeurIPS 2026 · similarity 0.87)
- Beyond Static Endpoints: Tool Programs as an Interface for Flexible Agentic Web Services (ICML 2026 · similarity 0.87)
- The Berkeley Function Calling Leaderboard (BFCL): From Tool Use to Agentic Evaluation of Large Language Models (ICML 2025 · similarity 0.87)
- ShortcutsBench: A Large-Scale Real-world Benchmark for API-based Agents (ICLR 2025 · similarity 0.86)
- NaviAgent: Graph‑Driven Bilevel Planning for Scalable Tool Orchestration (ICML 2026 · similarity 0.86)
- OSWorld-MCP: Benchmarking MCP Tool Invocation In Computer-Use Agents (ICLR 2026 · similarity 0.86)
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.
