Notes
🔍

Task Routing (Agent)

A large support team does not hand every ticket to the same person. Something reads the request first and decides who should take it — billing, returns, or a senior engineer. Task routing is that step inside a multi-agent system: a component inspects an incoming task and chooses which agent will do the work.

How the choice is made
The router sits in front of the specialists and decides on whatever signal is available:

  • Rules — keyword, file type or metadata matches; cheap and completely predictable;
  • Model-based classification — a small, fast model reads the task and names the agent best suited to it;
  • Declared capability — each agent advertises what it can do, as an Agent Card does under the Agent2Agent protocol, and the router matches the task against those declarations;
  • Cost and load — the routine case goes to a cheap model, and the expensive one is reserved for work that needs it.

A router that cannot decide confidently still needs somewhere to send the task: a general-purpose agent, or a person. Frameworks make this structural — a LangGraph conditional edge is routing written into the graph, and CrewAI expresses it as assignment across a crew.

Why it matters
Routing is what makes specialisation pay off. A misroute costs more than a wrong answer: the wrong agent works through the task with its own tools and budget, produces something plausible, and either hands it back or passes it on — and two agents handing the same task to each other will consume a whole run without finishing it.

Task routing is the step in a multi-agent system that decides which agent handles an incoming task, using rules, a classifier model, or the capabilities each agent declares. It is what makes specialisation worthwhile: work reaches the agent equipped with the right tools and permissions, routine cases avoid expensive models, and there is a defined fallback when no agent is a confident match. A misroute spends a full agent run on the wrong worker.

Phone a big company and the first person you reach often just works out who you should actually be talking to. They do not fix your problem — they know which department does.

Task routing is that receptionist, for a team of AI agents. When several agents each have their own speciality — one searches the web, one writes code, one handles refunds — something has to read the job and pick the right one. Choose well and the work goes to the agent equipped for it. Choose badly and an agent spends time and money on a task it was never suited to, then passes it along anyway.