Models, Rate Limits, and Spend
The scarce resource is frontier-model calls in a rate-limit window. Spend them where reasoning density is highest.
Antigravity's preview is free with rate limits that refresh in multi-hour windows — so the resource you're budgeting isn't dollars, it's frontier-model calls per window. The principle transfers to any paid setup later: match the model to the cognitive demand of the task, and cap what any one task can burn.
Default smart, downgrade deliberately
Frontier tier (Gemini 3 Pro-class) for planning, debugging, and design calls. Faster models for mechanical work — renames, doc updates, config chores. Set the model per agent when you spawn it.
Rule of thumb: if the task needs judgment, spend the frontier call. If it needs keystrokes, don't.VerifyAt least one of your running agents is on a cheaper model than your default — deliberately.Ride the window
If you hit limits mid-morning, don't stop working — shift mechanical tasks to lighter models and queue the judgment-heavy ones for the window reset.
Triage when limited: judgment tasks → wait for reset. Mechanical tasks → lighter model now.VerifyA rate-limit moment costs you sequencing, not progress.Cap blast radius with small briefs
A failed 10-minute task costs one retry. A failed 2-hour epic costs the afternoon. Brief size is your per-run spending cap — keep tasks small enough that a bad run is cheap.
Sizing check: if a run failing completely would cost more than 15 minutes of agent time, split the brief.VerifyYour Manager board shows several small tasks per repo, not one monolith.