AI Agent API Costs: Cut Spending 75% with Haiku Triage

The $340 Learning Curve

Spent $340 on API calls before my AI agent did anything useful.

Week 1: Everything on Opus

Opus for everything — $47 in 4 days.

Week 2: Tier the Models

Added Haiku for triage — 90% handled by cheap model, down to $12/day.

Week 3: Fix the Memory

Memory was 8-second latency on 50MB logs — switched to daily files + embeddings on summaries, now 200ms.

Week 4: OAuth Apocalypse

OAuth expired Saturday, agent retried 2,847 times, rate limited 24hrs — the same single-point failure that later took all 3 practical agent workflows down at once.

Current State

~$4/day, handles calendar/email/deploys, breaks something new every week. The gap between 'demo that impresses' and 'thing that runs' is massive.