AutoAgent
Type: full-code · Vendor: HKU Data Intelligence Lab · Language: Python · License: MIT · Status: active · Status in practice: experimental · First released: 2025-02-10
Allow non-coders to build and run LLM agents through natural-language dialogue — the framework profiles agents, generates tools and workflows, and runs them in Docker-isolated environments, with architecture drawn from OpenAI Swarm and Magentic-One.
Description. AutoAgent (HKUDS, formerly MetaChain, MIT) is a 'Fully-Automated & Zero-Code LLM Agent Framework'. It exposes three modes: User Mode (a ready-to-use multi-agent research assistant), Agent Editor (natural-language single-agent construction with automated profiling and tool generation), and Workflow Editor (natural-language multi-agent workflow construction without tool creation). Code generation and execution run inside Docker containers, with browser environment integration; design influences are explicitly OpenAI Swarm and Microsoft's Magentic-One three-agent pattern. Evaluated on GAIA and MultiHopRAG.
Agent loop shape. Three-mode shell. In User Mode a ready-built multi-agent research assistant runs over the user's task; in Agent Editor the user describes an agent in natural language and AutoAgent profiles it, generates required tools, and emits a runnable agent; in Workflow Editor the user describes a multi-agent workflow and AutoAgent synthesises an executable workflow without authoring new tools. Inside each agent the loop is tool-calling: model picks a function, the function runs inside a Docker container (file system, code execution, browser), the result is fed back. Architectural inspiration is OpenAI Swarm (handoffs) and Magentic-One (lead orchestrator + workers).
Primary use cases
- zero-code single-agent construction via natural language
- natural-language multi-agent workflow construction
- ready-to-use multi-agent research assistant (User Mode)
- GAIA-benchmark style autonomous problem solving
Key concepts
- User Mode (docs) — Ready-to-use multi-agent system for research tasks.
- Agent Editor — Natural-language single-agent construction: profiling + tool generation.
- Workflow Editor — Natural-language multi-agent workflow construction (no tool creation).
- Self-managing workflow generation → automatic-workflow-search — Dynamically creates, optimises, and adapts agent workflows from high-level task descriptions.
- Docker sandbox → code-execution — Code generation and execution run inside Docker containers.
- Swarm + Magentic-One inspiration — Acknowledged architectural influences for handoff-based and orchestrator-worker shapes.
Patterns this full-code implements —
- ·Automatic Workflow Search
Self-managing workflow generation: AutoAgent dynamically creates and adapts agent workflows from high-level task descriptions.
- ★Agent-as-Tool Embedding
Agent Editor produces agents that are themselves invoked as units inside larger workflows; tools are also LLM-generated artefacts.
- ★★Code Execution
Code generation and execution via Docker containerization.
- ★★Supervisor
Architectural inspiration explicitly includes Magentic-One's three-agent (lead-orchestrator + workers) design.
- ★★Tool Use
Function-calling tool use; LLM-generated tools are wired into agents at construction time; supports third-party tool platforms like RapidAPI.
- ★Browser Agent
Browser environment integration is part of the Docker sandbox setup; full GUI Agent / Computer-Use support is on the roadmap, not yet first-class.
- ★★Agentic RAG
Evaluated on MultiHopRAG via an 'Agentic-RAG Task' evaluation script; agentic retrieval is a benchmark capability rather than a core API.
- ★★Plan-and-Execute
Workflow generation produces an executable plan that subsequent agents execute; the workflow editor mode supports natural-language plan synthesis. AutoAgent does not use the 'plan-and-execute' label…
- ★★Structured Output
Agent profiling and workflow synthesis rely on structured outputs internally — automated profiling steps emit agent/workflow profiles. Not exposed as a user-facing schema API.
References
Provenance
- Last analyzed:
- Last updated:
- Verification status: verified