Runkle's rule for agents, simple by default, frontier by exception

Sydney Runkle compressed her own talk on building agents with decision models into five words, "simple by default, frontier by exception!!"

Sydney Runkle

@sydneyrunkle

X

simple by default, frontier by exception!!

Erin Fox, who was in the room, described the talk the same way. In her post she said she really enjoyed the talk on "Building Agents with Decision Models" and that the thinking is to start simple and be cheap by default. That is two accounts, speaker and attendee, landing on the same summary, which is about as much corroboration as a conference line gets.

Erin Fox

@erinfoox

X

the thinking is to start simple and be cheap by default

What is missing from it

The rule tells you the direction of travel and nothing about when to actually move. Neither post gives a threshold, an accuracy floor, a latency budget or a cost ratio that would tell you the point at which a cheap decision model stops being good enough and the frontier model earns its invoice. No eval, no workload, no numbers were posted with it.

So treat it as a default worth arguing with rather than a measured result. The version of this advice that fails in production is the one where nobody writes down the escalation condition, and the team either never escalates and ships a quietly wrong agent, or escalates on vibes and pays frontier prices on every call.

For people building on it, the useful move is to turn the slogan into something checkable before you adopt it. Decide what "needs it" means for your task, measure the cheap path against it, and only then decide what fraction of traffic goes up a tier. Runkle has posted ordering advice before, including her earlier sequence for fixing a misbehaving agent, which is similarly a priority list rather than a benchmark. Both are heuristics about where to spend your attention first, and both leave the measurement to you.