* [[Code Interpreter]] * Open Interpreter https://github.com/OpenInterpreter/open-interpreter * OpenAI O1 Code Interpreter https://github.com/FanaHOVA/openai-o1-code-interpreter * 只是簡單的實作 * CodeAgent: Enhancing Code Generation with Tool-Integrated Agent Systems for Real-World Repo-level Coding Challenges (2024/1) * https://arxiv.org/abs/2401.07339v2 * https://github.com/zkcpku/CodeAgent (404 了XD) * OpenCodeInterpreter: Integrating Code Generation with Execution and Refinement (2024/2) * https://arxiv.org/abs/2402.14658 * https://github.com/OpenCodeInterpreter/OpenCodeInterpreter * Executable Code Actions Elicit Better LLM Agents (2024/2) ⭐ * https://arxiv.org/abs/2402.01030 * https://github.com/xingyaoww/code-act * 核心 code: https://github.com/xingyaoww/code-act/blob/main/scripts/chat/demo.py * https://github.com/huggingface/smolagents 的 CodeAgent 也採用這個概念來實作 * 不採用 function calling API，而是直接在 generated code 裡面放 function 呼叫，然後整個 generated code 拿去執行 * 對比 Code Interpreter 只是當作一個工具來用，CodeAct 的概念是把 planning 跟 function calling 都合併在同一份 python code 一起執行，效率++ > 我用 Deep Research 研究報告: https://chatgpt.com/share/67a5e8ac-e820-8008-9329-225465aa63e0 * Code Mode: the better way to use MCP (2025/9/26) * https://blog.cloudflare.com/code-mode/ * Code Mode: give agents an entire API in 1,000 tokens (2026/2/20) * https://blog.cloudflare.com/code-mode-mcp/ * cloudflare 新的 MCP server，涵蓋整個 Cloudflare API * 它不再提供成千上萬個工具，而是只匯出兩個： `search()` 和 `execute()` * Claude: Give Claude a computer: PTC 功能 * https://x.com/rlancemartin/status/2027450018513490419 (2026/2/28) * claude api 的 `code_execution` tool 的「能力升級」， 從原本只能跑 Python 標準庫 / 內建套件，變成也能呼叫你自訂的 tools。