Routing is the new prompt engineering

Why the agent that picks up a task matters more than the prompt it receives, and how we score skills, tools and cost in under 40 ms.

Tomas Reyes

Chrome rail splitting into three paths, one lit green

Written by

Tomas Reyes

Reading time

7 min

3of 8 agents working

Most teams tune prompts for weeks and then hand every task to the same generalist agent. The output improves a little. The cost does not.

Scoring a task in 40 milliseconds

Eventide scores each task against the declared skills, tool access and permissions of every agent in the fleet, then weighs live load and cost ceilings. The winner gets the task with a context bundle attached.

What we learned

Specialists beat generalists on every benchmark we ran, as long as hand-offs carry memory. Without memory, specialisation doubled the rework.

Tomas Reyes

All posts