Cheaper runs and more dependable runs are the same discipline

The cost misconception

Most conversations about token cost in agentic systems start with the model — the price per token, the context budget, the choice of model. That is not where the cost actually comes from in a working system.

The cost comes from the knowledge layer: moving the wrong knowledge, re-reading it, and sending it in a shape the agent has to re-derive before it can act. A system that sends the whole document every time pays for the whole document every time, even if the agent only needs the one procedure, the one value, the one permission. A system that re-reads the same knowledge every turn pays for it every turn. A system that sends knowledge in a raw shape pays the model to do the work of turning it into something usable.

The model cost is real. It is not the main cost in a working agentic system, and fixing it by choosing a cheaper model does not fix the underlying problem, which is that the agent is paying for knowledge it should not be paying for.

The four token costs

These are the ordinary places token cost comes from in an agentic system. They are not exotic. They are the costs that show up when the knowledge layer is weak and the context is a bucket.

Grounding

Grounding is the discipline of tying an agent's claims and actions to knowledge it can point back to. When the agent says something or does something, grounding is what lets the company ask "what knowledge did you use, was it current, was it permitted" — and get an answer that is more than "the model said so."

Grounding is not a separate thing from token savings. It is the same discipline seen from the other side. The cheapest action is the one the agent does not have to re-read to justify. The most dependable action is the one the agent can show its basis for. The two are the same action, designed with the knowledge layer in mind.

The usual grounding mistakes are the same as the usual token mistakes, turned around:

• The agent acts on a context it cannot reconstruct later — so when asked for the basis, it has nothing to show.
• The agent acts on a version it cannot prove is current — so the basis is there, but the company cannot tell whether it was the right version.
• The agent acts on knowledge it could see but was not permitted to act on — so the basis is there, but the company cannot tell whether the agent was allowed to use it.
• The agent acts on a shape it had to re-derive — so the basis is there, but the company cannot tell whether the agent derived it correctly, because the derivation happened inside the model, not in the knowledge layer.

Grounding is what makes the agent auditable. Auditability is what makes the enterprise willing to let it act. The cheaper run and the more dependable run are the same run, designed with the knowledge layer in mind.

The relationship

The short version: token savings and grounding are the same discipline. Sending less of the right knowledge makes the agent cheaper. Making the basis retrievable makes the agent more dependable. Doing both is the same design decision — take the knowledge layer seriously at the point where it meets the agent.

This is the reason the token-savings and grounding argument is usually the one that lands first in a business setting. It is the argument that speaks the language the business already uses — cost and trust — and it is the argument that points back to the same source: the knowledge layer, designed as a delivery surface, with currentness, ownership, permission, deprecation, and grounding carried as properties of the knowledge, not as afterthoughts bolted on when the agent fails.

The model is still doing the reasoning. The knowledge layer is doing the rest. The cost and the trust both come from how seriously the knowledge layer is taken at the point where it meets the agent.

Where to start

If you are under pressure on cost and trust at the same time — which is the usual pressure in a working agentic system — the usual sequence is:

• Find the knowledge the agent is paying for but not using. The whole-document dumps, the re-reads every turn, the raw shapes the model has to re-derive. This is usually the biggest cost and the easiest to fix.
• Put currentness, ownership, and permission on the knowledge before it goes into the context. This is the fix that pays for itself in both cost and trust.
• Make the basis retrievable for the actions the agent takes. Grounding is the discipline that turns "cheaper" into "cheaper and more dependable."
• Stage the context for the action, not for the conversation. The agent's first look is the smallest surface that lets it decide what to do next. The rest comes in only when it is needed.

The goal is not to make the agent cheap. The goal is to make the agent cheap and dependable for the same reason — because the knowledge layer is taken seriously at the point where it meets the agent.