Amey-Thakur Agent specialist router
One agent covering many task types is mediocre at each, because the instructions that help one hurt another. A router classifies first and dispatches to a specialist, which improves quality and cost at the price of one extra step and a new failure mode: misrouting.
Method
- Define the categories by how work differs, not by topic. If two categories would take identical instructions, they are one category.
- Classify cheaply. The router does not need the strongest model; it needs a reliable one with a small, well-specified job.
- Give the router an explicit unknown. Forcing every input into a category is what produces confidently wrong routing. Unknown goes to a generalist or a human.
- Make each specialist genuinely specialised. Tailored instructions, examples, and tools. A router in front of identical agents adds latency and nothing else.
- Log the routing decision with the input. Misrouting is invisible otherwise, and the log is what lets you fix the categories (see agent-eval-design).
- Handle multi-category requests deliberately. Either split them or send to the broadest relevant specialist, decided in advance rather than per case.
- Let specialists reject. A specialist that recognises the work is not its own should return it rather than attempt it (see agent-handoff-protocol).
Boundaries
Routing improves fit; it adds a step that can itself be wrong, and a misroute is usually worse than a generalist attempt. It suits stable task distributions and needs recategorising as work changes. Maintaining many specialists costs more than one generalist, so the quality gain must justify it.