AutoGen

AutoGen is an open-source programming framework for agentic AI, originally released by Microsoft Research. Its central idea is the conversable agent: applications are built from multiple agents that exchange messages with one another, with tools, and optionally with a human in the loop. By framing automation as a conversation between specialized agents, AutoGen lets a developer compose patterns such as a planner agent delegating to a coder agent whose output is checked by a critic agent — division of labor, expressed as dialogue.

Capabilities

AutoGen supports large language model calls, tool and function use, code execution, and both fully autonomous and human-supervised workflows. It provides multi-agent conversation patterns for coordinating how agents take turns, when a conversation terminates, and how results are aggregated — from simple two-agent loops to group chats where a manager agent routes each turn to the most relevant specialist. The project offers Python tooling and a layered architecture, so builders can work at a high level with prebuilt agent types or drop down to lower-level messaging primitives when the prebuilt patterns do not fit. The framework's research pedigree shows in its flexibility: it has been widely used to prototype and study multi-agent designs before they hardened into products elsewhere.

Running it on your own models

Nothing in the multi-agent pattern requires a cloud API. Frameworks in this family speak to any OpenAI-compatible endpoint, which means an Ollama or llama.cpp server hosting an open-weight model on your own hardware can stand in for a hosted provider. That is the configuration that interests the sovereign operator: agent workflows over local inference keep your prompts, documents, and tool outputs on your own machines. Be realistic about the demands — multi-agent conversations multiply token consumption, since every inter-agent exchange is another model call, so throughput and context length on your local rig constrain how elaborate a crew you can field. And agents that execute code or touch tools deserve the same caution locally as anywhere: sandbox the execution environment and treat any untrusted text entering the conversation as a potential prompt injection vector, because an agent that acts on what it reads can be steered by what it reads.

Project status

As of 2026, Microsoft has placed AutoGen in maintenance mode, continuing critical bug fixes and security patches while directing new development toward a unified Microsoft Agent Framework that merges lessons from AutoGen and Semantic Kernel. Readers evaluating AutoGen today should weigh that trajectory against their needs: it remains widely studied, well documented, and a useful reference point for how multi-agent conversation frameworks are structured, but new greenfield projects may prefer an actively developed successor. We note this neutrally — in a fast-moving field, understanding the pattern outlasts any particular package, and AutoGen's conversable-agent abstraction has already shaped how the whole ecosystem thinks about agent orchestration.

When multi-agent is overkill

Honest guidance: most tasks do not need a crew. A single capable model with well-designed tools and a clear system prompt covers the bulk of practical automation, at a fraction of the token cost and with far simpler debugging. Multi-agent designs earn their overhead when roles genuinely conflict — a generator and an adversarial critic, a planner coordinating specialists with different tools — or when isolation between contexts is itself the feature. If you cannot articulate why two agents will outperform one agent with a better prompt, start with the better prompt; you can always promote a workflow to a conversation later.

For related orchestration tooling, D-Central also defines Semantic Kernel and the broader ReAct pattern that informs many agent loops, along with the system prompt conventions that give each agent its role. These entries describe each project on its own terms rather than ranking them.