Models and spend

Agents cost money to run, and the two levers that matter are which model drives each one and how much context it carries on every call.

Models are chosen per agent

The platform maintains a catalog of available models, and each agent is assigned one. A storefront answering simple questions and an agent doing multi-hour reconciliation don't need the same model — assigning them the same one is the most common way an agent platform becomes expensive.

Model changes are tracked per agent with a history, so you can see what an agent was running when it behaved well or badly. That matters when you're evaluating a downgrade: the question isn't "did it get cheaper" but "did the outcomes hold up".

Where the money actually goes

Two things dominate, and only one is obvious:

The model, per call. Straightforward, and the thing everyone looks at.

The context, on every single call. Less obvious and often larger. Every tool schema attached to an agent is re-sent on every model call it makes. A large third-party tool catalog wired to an agent that uses a handful of its tools can account for tens of thousands of tokens per call, replayed forever.

Descoping the tools an agent doesn't use is therefore a direct, permanent saving with no behavior change — usually the highest-leverage cost control available. See Capabilities and skills.

Bounding unattended work

An agent working a routine has no natural stopping point. Two ceilings:

• Per-routine spend caps bound what one standing job can cost per day.
• Approval thresholds force human sign-off above a limit you choose, routed through the escalation inbox.

Both are configuration, not agent etiquette.

Visibility

Spend is attributed per agent, and rolled up per organization. Usage is tracked alongside it, and costs are shown at fine precision rather than rounded to the nearest cent — small per-call numbers rounded aggressively stop being useful exactly when you're trying to compare two models.

Per agent you can see what it cost, over what period, and against what it delivered — which is the comparison that matters. An agent that costs more and produces better outcomes is a good agent. Cost alone doesn't tell you.

Practical advice

• Right-size per agent. Don't put your best model behind a FAQ.
• Descope aggressively. Unused tool schemas are pure recurring cost.
• Cap routines. Unattended work is where surprises compound.
• Judge by outcomes, not tokens. The ledger is the denominator.
• Watch model changes. If quality drops after a swap, the history tells you when it happened.