AutoAgent: Build a Running Multi-Agent System Without Writing Code

Most agent frameworks assume you're a developer. You write the orchestration logic — the code that coordinates which AI agents do what, in what order. You define which tools each agent calls. You wire the AI model calls together in code. AutoAgent was built for the person who knows what they want done but has no interest in writing that plumbing.

State a goal in natural language. The framework decomposes it into tasks, creates the agents that will handle them, and runs the workflow. That's the pitch, and it comes out of the Data Intelligence Lab at the University of Hong Kong, supported by a published paper (arXiv:2502.05957).

At a glance

What AutoAgent actually is

AutoAgent is a fully automated agent framework — meaning you don't define the agent structure, it does. You describe a goal, and the system figures out what agents are needed, creates them, and runs the workflow. The authors make a pointed observation in their paper: roughly 0.03% of the global population has the programming skills to build AI agents (systems that plan and execute multi-step tasks) using conventional frameworks. AutoAgent's entire design is a response to that gap.

The framework ships three interaction modes. User mode is a pre-built multi-agent system built for research, information retrieval, and report generation — you use it immediately, no configuration required. Agent Editor mode is a conversational interface where you describe a custom agent you want and the system builds it for you. Workflow Editor does the same for multi-agent pipelines. All three accept natural language as input. None require you to write code.

Under the hood, the architecture has four components: an agentic system layer that handles the operational infrastructure; an AI model execution engine that translates instructions into actions; a self-managing file system that handles state across the workflow; and a self-play customization module that lets agents refine themselves through iteration. The paper describes this combination as an "autonomous Agent Operating System" — an OS analogy, not just a framework.

What ships in the repo

The core deliverable is the three modes described above. On top of that:

• Model flexibility via LiteLLM — LiteLLM is a translation layer that lets you talk to any major AI model using the same interface. Swap between Claude, GPT-4o, Gemini, DeepSeek, Grok, Mistral, or Llama without changing your setup. The default is Claude 3.5 Sonnet.
• Deep research mode — a lightweight version of the multi-agent system optimized for information-gathering tasks. Comparable in scope to what OpenAI calls Deep Research, according to the paper's benchmark positioning.
• GAIA benchmark evaluation scripts — the team evaluated AutoAgent on the GAIA benchmark, a standard test for generalist multi-agent task performance. The paper reports it surpasses other open-source methods on that benchmark.
• Agentic-RAG evaluation — RAG (Retrieval-Augmented Generation) is the technique of feeding an AI relevant documents before it answers a question. AutoAgent includes evaluation tooling that tests how well the system reasons across multiple documents in sequence to reach a final answer.
• Docker-containerized agent environment — each agent execution runs inside a container, which keeps the host environment clean and makes the agent behavior reproducible.
• Self-developing workflows — the system can create, modify, and optimize its own workflows based on results, without waiting for manual intervention.

There are no formal versioned releases yet — the repo ships off the main branch.

Why this matters for non-coders

The persona this framework was built for is the domain expert who needs agents: a research analyst who wants a system that reads papers, synthesizes findings, and produces a structured report; a marketing director who wants multi-step competitive intelligence gathered and organized automatically; a founder who wants a workflow that monitors news, filters by relevance, and drafts a weekly summary.

Every one of those use cases can be wired together by a developer in CrewAI or LangGraph. AutoAgent is the version where you describe the workflow in plain English and it builds itself.

The practical test for whether User mode is enough: if you want to give the system a complex research question and get a structured report back, User mode handles that without any setup beyond API keys. If you want a custom workflow — "every day at 8am, pull these three sources, summarize by topic, and format for my Notion" — that's the Agent Editor or Workflow Editor, still no code, but requires spending some time in conversational configuration.

The Docker requirement deserves an honest note. Docker is not something most non-coders have installed. For a reader who has never used a terminal, that's a real barrier. The install steps are not difficult, but they assume comfort with a command line. This is not a click-to-install web app.

Where it fits

AutoAgent occupies a specific spot in a crowded field:

• CrewAI (50k+ stars) — developer-first, role-based multi-agent orchestration. Mature, production-grade, and widely deployed. You write Python to define agents and their roles. AutoAgent is what you reach for if writing Python is the barrier.
• AutoGen (42k+ stars, AutoGen 1.0 generally available in early 2026) — Microsoft's conversational agent framework. Strong developer tooling, growing no-code Studio interface. The Studio is a genuine no-code option, but it's hosted. AutoAgent is fully self-hosted and open.
• LangGraph — graph-based workflow framework, designed for stateful production pipelines. More control than AutoAgent, requires more development skill to wield it. Different audience.
• smolagents — Hugging Face's framework where agents write and execute Python as their primary action. Very different design philosophy; useful for code-first automation, not natural-language-first workflows.

The defining differentiator for AutoAgent is that the framework itself — the agent structure, the task decomposition, the multi-agent coordination — is generated from natural language. The others require a developer to define those things in code.

Verdict

Install, with a clear-eyed expectation about the Docker step.

AutoAgent is one of the most honest implementations of the "zero-code agent" idea: the framework does the structural work that a developer would otherwise do. The GAIA benchmark results and 9,500 GitHub stars from a February 2025 paper are meaningful signals that this is serious research being taken seriously by practitioners.

Two things to weigh before committing. First, no formal releases means you're running off the main branch — adequate for exploration and personal projects, less adequate for anything where stability matters. Second, Docker is a real prerequisite. The install is a 15-30 minute process, not a two-click setup. For a non-coder, that means either doing it yourself (possible, with the README) or asking an AI assistant like Claude Code to do it (the most practical path).

If you're an analyst, a researcher, or any knowledge worker who has stared at an agentic workflow and thought "someone else should be building this thing" — this is the project to try first.

For your agent to install this

You can skip this section — it's the install details your AI agent needs. If you're using Cowork, Claude Code, ChatGPT Work, Codex, or Hermes, say "install AutoAgent for me" and it will read the tables below.

System / OS requirements

API keys

Install commands

# 1. Clone the repo
git clone https://github.com/HKUDS/AutoAgent.git
cd AutoAgent

# 2. Install the package
pip install -e .

# 3. Copy and fill in the env file
cp .env.template .env
# Edit .env — add GITHUB_AI_TOKEN plus whichever LLM key you want to use

# 4. Start AutoAgent (default: Claude 3.5 Sonnet)
auto main

# Or: start with a specific model
COMPLETION_MODEL=gpt-4o auto main

# Or: start in deep-research mode only
auto deep-research

Docker must be installed and running before step 4. On macOS, install Docker Desktop from docker.com. On Linux, follow the official Docker Engine install guide.

Once running, the interface opens a start page with User mode (pre-built multi-agent research system) and Editor modes (Agent Editor and Workflow Editor) accessible via the navigation.