Routing agent work by cost¶
I treat model spend as a design input with a control on it.
Status: in daily use · checked Oct 26, 2026
In plain terms
Work goes to the cheapest model tier that can do it. A heavier tier is allowed, but only with a written reason, and some routes can't be reached at all.
The expensive failure in agent work is the right answer reached the costly way and repeated at machine speed. So cost is routed the way a risky change is: by a rule, in one place, with a reason on the record.
The tiers¶
| Need | Where it goes | What it has to show |
|---|---|---|
| Looking: where is X, how is Y wired | a scout on the cheapest capable tier, with no write access | nothing; this is the bulk of the traffic |
| Bounded change with an owned path and a command that proves it is done | the default coder tier | nothing; it is the default |
| Hard reasoning, planning, or work on a security boundary | a heavier reasoning tier | a written reason that names what the default tier can't do |
| Judgment, supervision, review of what would merge | a frontier model | the role itself; it decides, it doesn't type |
A tier is a level of model capability and cost.
The principle is the one from How I run AI coding agents: use the expensive model to decide and the cheap one to look. For the documentation refresh, read-only agents produced structured patches and a deterministic script applied them; coding workers continue to implement bounded software changes in isolated worktreesAn isolated git checkout for agent edits so the shared main checkout stays clean. Glossary.
Choosing the heavier tier is justified, not rationed¶
The heavier tier isn't scarce on purpose. The control is that the choice is explicit and someone wrote down why. The reason is checked against a minimum length by the same rule whether the choice arrives as a one-off dispatch or as a field in a multi-step plan, so there is one rule in one file and no second copy that can disagree. Words like "hard", "complex" or "needed" aren't reasons; the reason has to say what the cheaper tier can't do. A phase that keeps failing on the default tier doesn't earn the heavier one by failing: that is what the escalation step is for, and it costs nothing until it fires.
Routes that cannot be reached¶
- Priority-price variants. Some models come in a "fast" form that buys queue position at a surcharge, with the same model and the same answer. They are refused by name at every point where a model is resolved, and "fast" isn't a tier that can be selected.
- A model through another vendor's tool. Routing a model through a tool that bills it against a different quota drains that quota silently. It is refused wherever a worker is launched.
- A silent switch to a paid model. If the cheap tier is unavailable or its allowance is spent, the work stops and reports. It doesn't move to a dearer model on its own. A refusal is an outcome to read, not an error to route around (the decision).
One file decides which model a tier means¶
A single configuration file names the default model, the allowed set and the billing class of each. The scoutAn untracked, read-only investigation on the cheapest capable model tier. Glossary command, the dispatcher, the worker provider, the plan validator and the escalation step all read it, and a test fails if a new hard-coded model name appears in the tooling. Pinning a model literal in a plan is how a plan ends up running a model nobody maintains, so plans leave it unset and inherit the fleet default.
Cost is bounded where it is created¶
- Delegation. A worker may hand part of its job to child workers, but the ledgerThe run ledger: relational records of dispatched runs, with reapers and stall timers. Glossary reserves a cost cap for the whole tree when the first child is created, and refuses a request that would break it (lanes and the run ledger). Delegation is also rare in practice: only 30 of the 8,047 records in the run-ledger snapshot of Oct 26, 2026 are child runs.
- Fan-out. Read-heavy investigation is meant to avoid expensive subagents. A hook denies spawning one and points at the scout laneA dispatch pattern with a fixed contract: courier, phone line, scout, remote argv, or orchestrate. Glossary.
- Reviewers. The independent reviewerIn the main repositories, a separate agent whose recorded decision is a required pull-request check; only its identity can post it. Glossary's model is chosen per batch by how risky the change is: documentation and small configuration changes get the lighter reviewer, a standard change gets the same model working harder, and a change to a security boundary, a guard or a secrets path gets the strongest one. A reviewer can be raised by message, never lowered.
A cost lesson, learned by measuring¶
When a large documentation refresh was run the obvious way, the per-agent usage was measured halfway through instead of guessed, and it showed long context being re-read on every turn. The run was stopped, the good work was kept, and the bulk of it moved to read-only investigators on the cheap tier that returned structured patches for a script to apply. The full story is in How I run AI coding agents.
Beyond hosted models¶
GPU compute on demand: designed, parts built
Besides hosted models, I have designed a way to serve open-weight models from a rented GPU that exists only while someone wants it, so it costs nothing while idle. Parts of it are built.
Where a measurement exists, it is printed with its window, as in measuring agent reliability.