A R$0-token day of agent work — the routing ladder, and the two warnings nobody gives

2026-08-29 · method note · numbers here are stated with their premises, not benchmarks

Yesterday a full workday of agent tasks — reading, summarizing, triaging piles of material — cost R$0 in API tokens. Not a promo credit, not a free tier trick. One architectural decision: routing.

The common mistake: one model for everything

Most people use AI like a single brilliant employee: throw everything at the most impressive model. But roughly 80% of an agent workday's volume isn't hard. Reading, summarizing, classifying, extracting fields, formatting. Paying frontier-model prices for mechanical work is hiring a surgeon to change a lightbulb — he'll do it, and you'll pay his full rate.

The ladder

Give each task the lowest rung that can actually do it:

• Rung 1 — raw volume (harvesting, extraction, mass classification): small local model, running on your own machine. Token cost: zero.
• Rung 2 — reading and synthesis (long material, comparing sources): mid-size local model. Still zero.
• Rung 3 — hard decisions (final verdicts, judgment-heavy analysis, critical code): the frontier API, and only there. You pay for the rung, not the ladder.

Who picks the rung? A simple classification rule is enough to start — or the local model itself labels the task before executing it.

What a R$0 day actually is

A zero-token day is a volume day: everything landed on rungs 1 and 2. The local model ate the material, agents logged everything to files, results came back ready for review. And the honest number for energy: under R$1 per long cycle on typical hardware — premises vary with your machine and tariff. R$0 in tokens isn't R$0 absolutely; it's trading per-request cost for the fixed cost of a machine you already own.

The two warnings gurus skip

1. Cheap that redoes the work three times is expensive. If the small model fails and retries, your cost per completed task climbs — price-per-token is the wrong metric to decide with. Measure cost per completed task, always.

2. Routing quality down without a reviewer manufactures error in silence. Every step down the ladder needs quality sampling: periodically, a human or a second agent checks a sample of what the cheap model produced. And some things never go to rung 1 — final decisions, anything published under your name, anything you can't verify afterward.

Three rules to keep

• Mechanical and voluminous → local model.
• Hard decision → frontier API, only there — it's worth paying, because it's the price of being wrong less.
• Always measure cost per completed task, never per token.

Questions about where routing breaks in practice are welcome — that's the interesting part.