Agentic AI · Part 6 of 7
Multiple Agents and Frameworks
Routing between several agents, and the frameworks that orchestrate them.
More than one agent
The same loop extends to several agents. A common next step is a supervisor: an LLM call whose only job is reading the current state of a conversation and deciding which of several specialized agents should act next, or whether none of them have anything useful to add, in which case it ends the round. Each agent it routes to can have its own system prompt, its own tools, even its own private data the others can't see.
Anthropic describes a related pattern, orchestrator-workers: a central model breaks a task into subtasks, delegates them to worker models, and synthesizes their results, with the subtasks determined at runtime. Giving each worker its own context window is also a way to manage context, as in Memory and Context.
This site's own Roundtable project is a live example: sign in and the app generates a mystery on the spot, then you pick up to three companions, from a roster of eight, each an expert in their own fictional universe, to investigate it with you. A supervisor node decides which of your chosen companions speaks next each round. How many speak is set differently: a random target count, from one up to the number of companions you picked, is fixed when the round starts, because letting the supervisor decide when to stop was unreliable in practice. Each companion retrieves from its own private trove of lore before replying, the same tool-calling pattern as in Tools and Tool Calling with a vector search in place of a calculator.
Frameworks
Hand-rolling the loop works, but it gets repetitive once there's more than one agent: routing between them, retrying a tool call the model formatted wrong, keeping track of whose turn it is, streaming partial output to a user. Frameworks handle that plumbing:
- LangGraph (LangChain) is a low-level orchestration framework and runtime for long-running, stateful agents. You build a graph in which each agent, tool call, or routing decision is a node and edges control what runs next. Roundtable is a LangGraph graph with a supervisor node.
- AutoGen, from Microsoft, was an early and widely used framework built around agents conversing in a shared “group chat.” It is now in maintenance mode, community managed, and its README points new users to Microsoft Agent Framework.
- CrewAI is built around role-based agents grouped into “crews” that work on assigned tasks. Its docs recommend wrapping crews in a “Flow” that controls state and execution order, and calling a crew from a Flow step when a task needs a team.
- Google's Agent Development Kit (ADK) supports the Agent2Agent (A2A) protocol, which lets agents hosted as separate services communicate (marked experimental in ADK's docs).
- Microsoft Agent Framework is the direct successor to AutoGen and Semantic Kernel. It supports Python and .NET (Go is in public preview), offers both agents and graph-based workflows, and works with Microsoft Foundry (formerly Azure AI Foundry) and other model providers.
All of them wrap the pieces this series covered: the loop, tools to call, and, with more than one agent, a way to decide who acts next. They differ in how explicitly they model that control flow.