A year running a multi-model router: 5 design choices that survived, 3 that died

AIGODSEND · reddit · 2026-10-19

The author ran a drafter/critic/long-context-search three-model router behind one interface, with a provenance-tagged lessons file and a GC that deletes any lesson uncited for 30 days — for about a year of real work.

What survived:

Roles, not personalities: the critic only reads failures, the drafter never sees its own wins; separating memory scopes beats naming agents.
Verification gates: nothing exits without a checker (tests, a second model, or a human) — "an unchecked model output is just a rumor with good formatting."
The lessons file: every failure becomes a rule with date and provenance; the GC is the hardest worker in the house.
Small hot context: 64k resident, everything else searched and cited, never stuffed.
Temperature per task: 0.2 extraction, 0.7 drafting, 0.9 brainstorming with a human picking survivors.

What died:

Agent names and personas — cute for a week, useless after.
Voting between models — agreement is peer pressure; models share failure modes from the same internet.
Autonomous chains over three steps — compounding failures, unreadable audit trail; checkpoint every three steps.

Funniest failure: the critic eventually rejected everything and had to be given a quota of at least one approved solution per day.

Original post →

A year running a multi-model router: 5 design choices that survived, 3 that died

More from coding & agent