The Most Expensive Employee Is the One That Does Nothing

The 2026 conversation about agent cost is loud and almost entirely aimed at the wrong target. The dominant frame is cost per token, and the dominant advice is to trim tokens — shorter prompts, cheaper models, caching, compression. All useful. None of it is the interesting decision.

The interesting decision shows up the moment you build an agent that runs all the time.

Two very different economic shapes

Most agents people ship are serve-on-demand: a user asks, the agent works, the agent stops. A chat feature, a one-shot report generator, a coding assistant. The cost scales with usage, so cost-per-token is a perfectly fine proxy. More users, more tokens, more bill. It behaves like the SaaS you already know.

An always-on agent is a different animal. It watches continuously, decides almost everything is irrelevant, and acts on a tiny sliver. Traffic Driver monitors twelve platforms, and the overwhelming majority of what it sees gets discarded before any human would ever know it existed. The cost does not scale with value delivered. It scales with time.

That breaks the mental model. An agent that does nothing ninety-nine percent of the time is, under the cost-per-token lens, the most expensive employee in the company — it’s being paid to sit in a chair. And an entire class of “is this agent even worth running?” panic comes from measuring it with the wrong yardstick.

The number that actually decides viability

For a hunt-then-act agent, the metric that matters is cost per decision — total spend divided by the decisions that mattered. Not per token, not per run, not per day. Per decision. A close second is cost per caught event: how much did it cost in idle inference to surface the one real opportunity or threat that made the agent worth having at all.

Those two numbers behave nothing like cost-per-token. Caching and compression barely move them, because the spend isn’t in the interesting runs — it’s in the ten thousand boring ones that produced nothing. The lever that moves them is the watch loop itself: how cheaply can the agent look at something and correctly decide to do nothing.

Which reframes the whole build. The cheapest always-on agent isn’t the one with the cheapest model. It’s the one with the cheapest “ignore this” — the cheapest reliable way to filter the noise before expensive reasoning ever touches it.

The false-positive tax

There’s a trap hiding in the above. The cheapest possible ignore is “ignore everything,” which scores wonderfully on cost-per-decision and catastrophically on the metric you actually care about: did you catch the thing? So the real engineering is balancing two opposing costs — the cost of looking and the cost of missing — and every product has a different exchange rate.

For Traffic Driver, a false positive means a weird post going out on a client’s account. For Proven Duty, a missed event is a compliance gap a regulator could find. The exchange rate between “look cheaper” and “miss less” is the entire product decision, and it’s a product decision dressed up as a DevOps line item.

Why this reframes build vs. buy

This is exactly why I keep reaching for an off-the-shelf evaluation and observability layer every time I start a new agent, and why I care that it actually fits the always-on shape. The work that decides whether a monitor is affordable — cheap filtering, scored confidence, durable traces, cost tied to decisions rather than runs — is the work a proper layer is supposed to absorb. The model vendors sell you cheaper tokens. Almost nobody sells you cheaper looking.

Until that layer exists and fits, the agents that survive in production are the ones whose builders understood the economics of standing still. The ones that don’t are the projects in the forty-percent-canceled bucket — not because the agent couldn’t do the job, but because nobody measured the right cost and it quietly ate its own budget.

---

An agent’s most expensive state is the one where it’s right to do nothing. Price that correctly, or price yourself out of running it at all.