The Dispatcher

Routing AI work to the cheapest thing that can do it.

Frontier model spend scales with carelessness,not with actual need.The bill arrives later.

02 / PROBLEM

The failure mode

Frontier model spend scales with carelessness, not with actual need. Once agents are making dozens of calls a session on your behalf, “just use the best model for everything” stops being generous and starts being a slow-motion bill you don’t see coming.

03 / CONSTRAINTS

Operating envelope

The policy had to be something a human can reason about and audit, not an opaque auto-router nobody trusts enough to leave alone.
Free local compute exists, but it’s meaningfully weaker — routing everything there tanks quality, so it’s not a universal answer.
Deciding “which tier does this task deserve” has to itself be cheap, or the routing overhead eats the savings it’s supposed to create.
And a written policy nobody actually follows under deadline pressure isn’t a policy, it’s a wish.

04 / THE SYSTEM

How it holds together

→ lateral reasoning effort ↗ vertical price escalation ⋔ downward free fanout

Read the system narrative

A tiered roster: a floor model handling the large majority of work, a mid-tier escalation reserved for architecture and ambiguous judgment calls, an apex tier reserved for genuine cross-domain adjudication, and a fast/cheap tier for pure high-volume fanout. Two levers, used in strict order: raise reasoning effort within the current tier before jumping to a pricier one — most “I need a smarter model” moments turn out to be “I need this model to think harder,” which costs nothing extra. A “native advisor” pattern pairs a cheap executor with a single read-only consult from a pricier model, so a task gets most of the expensive tier’s judgment quality without running the whole thing on the expensive tier. Local model offloading soaks up commodity generation — boilerplate, docstrings, scaffolding — so cloud calls are reserved for things that actually need judgment. And underneath all of it, a pre-execution gate blocks a task from silently running on the wrong tier, rather than trusting agents to self-police a policy they didn’t write.

05 / WHAT BROKE

INC-20260708

Incident replay

Symptom: a routine cost review turned up a spike traced back to small fanout tasks.
Root cause: background subagents were silently inheriting their parent task’s model tier by default — invisible for weeks, technically documented, practically ignored — dozens of individually-trivial tasks quietly running at the top tier the entire time because nobody had made the tier explicit.
Fix: model tier became an explicit, required, machine-enforced field.

SYMPTOM: a routine cost review turned up a spike traced back to small fanout tasks. ROOT CAUSE: background subagents were silently inheriting their parent task’s model tier by default — invisible for weeks, technically documented, practically ignored — dozens of individually-trivial tasks quietly running at the top tier the entire time because nobody had made the tier explicit. FIX: model tier became an explicit, required, machine-enforced field. Separately: an early routing classifier occasionally misjudged a genuinely hard task as commodity work and shipped a subtly wrong answer with total confidence — worse than an expensive right answer — so classifier-only routing got scrapped in favor of the explicit-lever model above.

06 / RETROSPECTIVE

What I’d do differently

Build the enforcement gate before writing the routing policy, not after the audit that caught the leak. A policy that isn’t machine-enforced is a suggestion, and suggestions erode under deadline pressure exactly when discipline matters most.

07 / SPEC PLATE

Build record

Status: live LIVE, actively tuned
Stack: multi-tier model roster, local model mesh, enforcement gate.
Scars: the silent inheritance spike.
Last incident: 2026-07-08