AG2 (formerly AutoGen)
Status:
assess| Maturity:emerging| Last reviewed: 2026-03-20
What it does
AG2 (formerly AutoGen) is an open-source Python framework for building multi-agent AI systems. It allows developers to define multiple LLM-powered agents that communicate with each other to solve complex tasks through structured conversation and tool use. Originally developed by Microsoft Research, it was forked and rebranded as AG2 in 2024 by a community-led group.
AG2 is used for: complex task decomposition (splitting a large task across specialist agents), automated code execution pipelines, research workflow automation, and building custom agentic applications.
Key capabilities
- Multi-agent orchestration: define agent roles, communication patterns, and workflows
- Human-in-the-loop support: pause agent workflows for human review/input
- Tool use: agents can execute code, call APIs, query databases
- GroupChat: multiple agents collaborate on a task via structured conversation
- Support for multiple LLM backends (OpenAI, Anthropic, Gemini, local models)
- Teachable agents: agents can learn from interactions
- Nested agents and sub-workflows
Fit for Sheffield researchers
AG2 is a developer framework — relevant to research software engineers, computational researchers, and CiCS staff building automated research pipelines or AI-powered research tools.
Potential use cases at Sheffield:
- Automated data collection and preprocessing pipelines
- Multi-step literature review automation
- Autonomous laboratory experiment scheduling (in combination with lab hardware interfaces)
- Research administration workflow automation
Not suitable as a direct researcher self-service tool — requires Python development skills.
Licensing and access
Apache 2.0 open source. Install via pip (pip install ag2). No direct cost for the framework; LLM API costs depend on providers configured.
Trust and governance notes
AG2 itself does not handle data — data governance depends entirely on the LLM APIs configured within agent workflows. Researchers building AG2 pipelines must ensure that any data passed to external LLM APIs (OpenAI, Anthropic, etc.) is handled in accordance with data processing agreements and Sheffield’s data governance policies.
For sensitive data workflows, configure AG2 to use locally-hosted LLMs (via Ollama or vLLM) to keep data on-premise.
Agentic systems with autonomous tool execution carry higher risk than single-step AI interactions — all code execution agents should operate in sandboxed environments.
Observed limitations
- High technical barrier — requires Python and LLM API expertise
- Multi-agent systems are complex to debug and reason about
- Rapid framework evolution (AutoGen → AG2 fork) creates API instability
- Agentic execution errors can compound — robust error handling and checkpointing needed
- LLM API costs can be significant for complex multi-agent workflows
- Limited out-of-the-box integration with scientific tooling
Related tools
claude-code · litellm · chemcrow · labos
Sources and evidence
- AG2 website: https://ag2.ai (accessed 2026-03-20)
- AG2 GitHub repository: https://github.com/ag2ai/ag2 (accessed 2026-03-20)
- AG2 documentation: https://docs.ag2.ai (accessed 2026-03-20)
- Original AutoGen paper: Wu et al., 2023: https://arxiv.org/abs/2308.08155 (accessed 2026-03-20)