Agent orchestration

1 · What it is

Some AI apps use several agents. An agent here is a language model that can use tools and take steps on its own. Agent orchestration answers three questions. Which agents work? In what order? And who chooses what happens next? OpenAI’s Agents SDK (a toolkit for building agents) names two ways to do this. Either the model decides, or your code decides. Anthropic draws a similar line. In a workflow, code that was written in advance sets the path for models and tools. In an agent, the model steers its own steps and picks its own tools. LangGraph’s docs agree that workflows follow paths fixed ahead of time, in a set order.

When the model decides, there are two common moves. The first is a handoff. A triage agent, which sorts incoming requests like a front desk, passes the chat to a specialist. The specialist then takes over. The second move keeps a lead agent in charge. It calls specialists like tools, each for one small job. You can mix both moves. Microsoft AutoGen has a group chat where a model reads the shared conversation. It then picks which agent speaks next, using each agent’s name and short description. By default it will not pick the same agent twice in a row. AutoGen’s example team does web search and data analysis.

When code decides, the path is written down before anything runs. Google’s Agent Development Kit (ADK) has ready-made workflows. They run agents one after another, side by side, or in a loop. No AI model is asked to pick the order. In ADK’s side-by-side example, three research agents each save their result in a shared store. A merger agent then combines them. LangGraph can also create worker steps as it goes, each with its own working memory. The workers write their results to a shared place. The orchestrator, the part in charge, reads them to build the final answer.