adobe-research/dynasaur

Official repository for "DynaSaur: Large Language Agents Beyond Predefined Actions"

DynaSaur is a framework for building AI agents that can dynamically create and reuse actions as Python code. Instead of being limited to a fixed set of predefined tools, the agent writes new Python snippets when needed—either from scratch or by combining existing actions—to solve tasks. These newly created actions are saved to a growing library for future use, enabling the agent to adapt and improve over time. The system is designed to handle unexpected situations by generating appropriate responses even when no relevant actions exist initially. It has achieved top performance on the GAIA benchmark, a challenging test of general AI assistant capabilities.

Related

  • Project
    OpenHands/software-agent-sdk

    A developer SDK providing APIs to build AI agents that can write code, manage files, and execute terminal commands in local or ephemeral workspaces.

  • Project
    Upsonic/Upsonic

    A Python framework for building autonomous and traditional AI agents, featuring secure workspace execution, custom tool integration, and a unified OCR pipeline.

  • Project
    MervinPraison/PraisonAI

    PraisonAI is a Python‑first framework for building autonomous AI agents with a five‑layer stack (Prompt, Context, Harness, Loop, Graph) plus managed remote runtimes. It offers SDKs, CLI, UI/dashboard, JavaScript support, and works with 100+ LLM providers for use‑cases like research, coding, content, data pipelines, support bots, and workflow automation.

  • Project
    DannyMac180/skills

    A GitHub repo offering plug‑in style “skills” that extend Claude Code or any skill‑aware AI agent with ready‑made ModSmith mods, multi‑agent workflow orchestration, task relaying, prompt‑ambition boosting, and artifact explanation with spaced‑repetition learning.