Mikie vs Augment Code: closest thesis.
Augment proved token-efficient context wins. Mikie is pushing the same idea harder — explicit savings floors, failure-taxonomy routing, and cost per resolved task.
What Auggie is
Auggie is Augment's coding agent. Their published Terminal-Bench 2.0 run (Opus 4.7, five attempts per task, Harbor) is the closest public comparison surface to what Mikie measures.
Auggie: Augment's published Opus 4.7 run. Mikie: local Terminal-Bench 2.0 full run, v2.8 path, GPT-5.5 low reasoning (58/89 completed scored). Different models, attempt counts, and harness settings — NOT an apples-to-apples public claim.
The honest read
Auggie is ahead on published pass rate, and we say so. Mikie's local run is 91% cheaper and 84% lighter on tokens, but that stays a local measurement — not a leaderboard claim — until a matched official submission clears our own claim gate.
Verdict
Same thesis, different discipline: Mikie ships the measurement ledger and failure taxonomy so every number above traces to a row you can audit.
Measure it on your own repo.
Every claim above traces to a measurement row or is labeled as positioning. That’s the product.