Cost-aware AI engineering: keeping agent spend visible and bounded

A single AI coding agent’s token spend is easy to eyeball — you’re watching the session, you have a rough sense of how long it ran. That intuition stops working the moment a workspace has several agents running concurrently across several tasks. Spend that used to be “roughly what I’d expect” turns into a number you only find out about at the end of the month, from a bill that doesn’t break down by task, by agent, or by model. That’s not a cost problem so much as a visibility problem — and visibility is fixable.

The number that actually bills, per task

CodeHerder tracks cost at the task level, not just the workspace level. Every task shows its own running cost in plain sight — the real number, tied to the actual work that produced it, not an aggregate total that hides which piece of work was expensive and which wasn’t. You can see cost broken down by agent and by model, over whatever window you care about, from the CLI as easily as from the dashboard:

ch costs workspace --window week --by agent,model

That’s not a summary generated after the fact — it’s the same data the dashboard renders, available to your own scripts and CI the moment you want it.

Nobody’s key runs through us

Cost-awareness only means something if the number you’re looking at is the real one. CodeHerder never stores your Anthropic API key and never bills you for token spend — you bring your own key, it stays configured on your own machines, and your token usage goes straight to your provider’s dashboard, not through a CodeHerder markup. What you pay CodeHerder for is the coordination layer: routing, workflow, memory, and visibility. What you pay your model provider for is tokens, directly, at their price.

Concurrency, not seats, is what’s metered

Plans are metered on concurrent agents — the work actually in flight at any given moment — not on how many agents you’ve registered. An agent sitting idle in a queue, parked in a done state, or simply registered but not yet claimed doesn’t use a slot or cost you anything extra. The moment a task finishes, that slot frees immediately for the next one. That means the cost model matches the thing you actually care about controlling: how much is running right now, not how many agents exist on paper.

Bounded spend starts with bounded surprise

None of this requires trusting a forecast. It requires the actual number, attributed correctly, available whenever you ask for it — per task, per agent, per model, this week or this quarter. Cost-aware engineering isn’t a discipline you impose after the fact; it’s what happens automatically once the number stops being a mystery.