Most AI Work Can Wait

Most teams building agents pick the model first & the architecture second. That is backwards. The model choice is the last decision, not the first.

What matters is the router, a small piece of code that decides which tier of model handles each request. Get the router right & 70-80% of traffic runs on local models that cost nothing per call, or on async models 1 that reduce AI spend by 90%+.

Brian Armstrong made the same point last week about how Coinbase cut AI spend in half while token usage grew 2 , paraphrasing :

How to keep AI spend flat while token usage grows exponentially : not with friction & spend alerts. With better defaults, routing, & caching. Engineers can choose any model they want, but defaults matter.

The routing problem has three layers, and each does a distinct job :

Skill classifier turns a raw user request into a concrete operation. It answers what the task is. Draft-a-reply, summarize-a-repo, run-a-migration. The classifier is intent recognition.

Router decides which tier executes the classified operation. It answers which model runs it. The router does not read the prompt. It reads the classifier’s label plus a few features : complexity, context size, historical success rate.

Model selector picks the cheapest model within a tier that meets a confidence threshold.

Classifier & router are not the same. The classifier is a language problem ; the router is a scheduling problem. Conflating them buries the model choice inside the prompt & kills the ability to A/B different models against the same operation.

Local compute is close to free. Async batch reasoning runs two orders of magnitude cheaper than real-time inference 1 . So the real question is narrower : what fraction of work needs real-time answers?

Surprisingly little, once the system can queue work.

Queueing is why this works. A draft reply, a repo summary, a diligence memo, a nightly evaluator run : none of these need to return in a second.

We built the first version of this into our agent runtime. The router already scored tasks on complexity, context size, & local memory retrieval. Two feedback mechanisms now sit on top of the router, & they operate on different time scales :

Synchronous failure-mode signals. A predictor annotates each incoming route with five features : missing repo context, long dependency chains, risky migrations, security-sensitive prompts, & high-consequence writes.

Nightly closed-loop feedback. A batch evaluator scores yesterday’s traces overnight & updates the router’s weights, running on async inference on Sail to keep the evaluation cost near zero.

The synchronous predictor catches known-hard tasks before they fail. The nightly loop discovers new failure modes the predictor missed.

Once skill distillation flattens the operation set, 70-80% of agent traffic can run on local models 3 for most non-coding work.

The implication : design your system around routing, not around models. Pick your models last.

Full Sail on Asynchronous Inference — the cost delta between real-time & async batch inference. ↩︎ ↩︎

Brian Armstrong on X — Coinbase cut AI spend nearly in half while token usage grew, via better defaults, routing, & caching. ↩︎

Skill Distillation , Teaching Local Models to Call Tools Like Claude , & The Minimill of AI . ↩︎