You Hired a Million Bad Employees. Give Them OKRs and a Budget.

One team cut an AI agent’s token consumption by 94% and its latency by 87%, and the change was not a smarter prompt. Vivek Haldar took an agent skill whose stable steps (fetch, filter, inventory) ran through the model on every invocation, and compiled those steps to Python. The model now runs only where judgment is required. Everything deterministic left the token budget entirely.

Hold that number against the one Adam Mosseri floated this week. Meta’s head of Instagram told an interviewer that within one to two years, a strong engineer’s token burn rate “might be the same as their salary,” and that companies will start setting token budgets per engineer, calibrated to how much they trust that person’s return on the spend. Two numbers, same lesson. When an agent’s work is undefined, tokens flow without producing anything, and someone eventually pays the bill.

The bill is already arriving

The token economy stopped being theoretical this quarter. Uber exhausted its entire 2026 AI coding budget by April. Microsoft discontinued Claude Code licenses over cost. Meta killed its internal token leaderboard once the projected spend crossed into the billions. We wrote about that leaderboard’s death in why usage is not impact: counting tokens rewards the wrong behavior, so Meta stopped counting them as a trophy.

Killing the leaderboard was the first move. It removed the incentive to burn. It did not answer the next question, which is what you replace the trophy with. Mosseri’s answer is a budget: a per-engineer cap tied to trust in ROI. That is a real governance instrument, and it exposes the mechanism underneath the whole problem.

Andreessen Horowitz named the mechanism in a piece titled, aptly, “You Just Hired a Million Bad Employees.” Their estimate is that 80% of tokens accomplish nothing. Agents loop. They re-read context they already have, re-derive conclusions they already reached, retry paths that already failed. As a16z put it, “you are spending tokens on spending tokens.” A million eager workers who never tire, never escalate, and never notice they are stuck. Cheap tokens are not cheap labor. They are labor with no job description, and unmanaged labor is the most expensive kind.

Manage agents as headcount

The framing that survives contact with the invoice is headcount. You already know how to manage a workforce whose cost scales with activity, because you have managed humans. A human employee has a role, an objective, a review cycle, and a manager who notices when they spin. Agents have none of that by default. They have a prompt and an API key.

Give them the rest. The objective an agent needs is not a mission statement, it is an eval. An eval defines what a correct output looks like, scores every run against that definition, and fails loudly when the agent drifts. In headcount terms, evals are the agent’s OKRs. They are the difference between an employee who knows the target and one who works hard in a random direction.

This is where per-engineer budgets and evals meet. Mosseri’s cap decides how many tokens a person’s agents may burn. The eval decides whether those tokens bought anything. Per OpenAI’s recent announcement on measuring the business value of agents, the metric that matters is useful work per dollar, not the price of tokens. A cap without an eval is a spending limit on activity you cannot grade. An eval without a cap is a quality bar with no cost ceiling. You need both, and they are the same tools every operations leader already uses on human teams: a budget and a definition of done.

Compile the loops out

Budgets and evals govern the agent. They do not shrink the work the agent should never have been doing. That is the second half, and it is the one most teams skip because it feels like premature optimization. It is not. It is the highest-leverage move available.

Look at what an agent actually does across a hundred runs. A large share of its steps are stable and repetitive: the same API call, the same filter, the same lookup, the same format conversion. Those steps do not need a language model. They need a function. Every time a deterministic step runs through the model, you pay full token price for something a compiler would do for free, and you inherit the model’s latency and its small but nonzero chance of getting a mechanical task wrong.

Haldar’s 94% came from exactly this discipline. He audited the skill, separated the steps that require judgment from the steps that are mechanical, and moved the mechanical ones into code. The model stayed only where it added value, at the decision points. The result was cheaper, faster, and more reliable at once, because deterministic code does not hallucinate a filename.

This is the operations tax we described in running AI at scale, paid down at the source. Every stable step you compile out of the model is a step that never appears on a future invoice, never contributes to a per-engineer cap, and never loops. The a16z 80% is not a fixed cost of doing business. A meaningful slice of it is undefined work that should have been code.

The shape of the fix

Put the three moves in order and they form one operating model. Cap the spend, so token consumption becomes a managed line rather than an open tab. Define the evals, so every capped token is graded on useful work. Compile the stable steps out, so the model only runs where judgment lives. Caps came into focus this quarter because the bills came due; the flat-fee era of unmetered AI is ending, as we argued in cost governance after flat fees. Budgets are the response, and evals plus compilation are what make a budget survivable instead of merely restrictive.

The organizations that will operate agents profitably over the next two years are not the ones running the most agents or burning the most tokens. They are the ones treating agent spend the way they treat payroll: a budget per owner, a defined objective per role, and a standing effort to move repetitive work off the expensive path.

Do this now

Pick one agent in production and run the headcount review. Pull its token consumption for the last thirty days and divide by useful outputs; that ratio is your cost per unit of work, and it is probably worse than you assume. Then read a sample of its runs and mark each step as judgment or mechanical. Every mechanical step is a candidate to compile into code. Set a monthly token cap for the agent’s owner, and attach one eval that scores whether the output was correct. You now have a budget, an OKR, and a shorter path. Do it for one agent this week before your CFO does it for all of them.

---

This analysis synthesizes You Just Hired a Million Bad Employees (Andreessen Horowitz, July 2026), Meta’s Adam Mosseri says AI token budgets could soon be capped per engineer (TechCrunch, July 2026), How I Cut an AI Agent’s Token Use by 94% (Vivek Haldar, July 2026), and OpenAI’s announcement on measuring the business value of AI agents (July 2026).

Victorino Group helps enterprises turn agent spend into a managed budget line with defined evals. Let’s talk.