Every runaway agent story starts the same way. The prompt got engineered. The tools got wired. The demo landed and the room bought in. Then it shipped, and the bill arrived in a shape nobody expected.
Here is the part we keep skipping: the agent had a personality but no wallet. It could reason, plan, and call tools, but nothing told it when to stop spending. That is a design gap, not a monitoring gap.
TechCrunch put numbers on the scramble. Expenditure on AI is exploding in large part due to agentic features, with per-developer consumption rising about 18.6x in nine months. And the companies feeling it hardest are not the reckless ones. Companies that gorged themselves in early 2025 on all-you-can-eat subscriptions are now scrambling to understand where their money is going, pull back spending, and figure out whether they can salvage some ROI from the wreckage of their budgets.
Cheaper tokens, bigger bills
The trap is the price curve. Everyone sees token prices falling and assumes the cost problem solves itself. It does not, because agents do not make one call. Agents don’t make one call. They loop, plan, call tools, retry, hand off, and resend their full context at every step. By step 20 you’ve paid for the same history 20 times.
That architecture is the multiplier. Agentic workflows consume 5 to 30 times more tokens per task than a standard chatbot query. Multiply that across a fleet and the math turns hostile fast. Even a well run organization gets blindsided.
Uber is the case study nobody wants to be. Claude Code adoption jumped from 32% to 84% of Uber’s 5,000-engineer org between December 2025 and March 2026. By April, the entire annual AI budget was gone. Their CTO said the quiet part out loud: “I’m back to the drawing board, because the budget I thought I would need is blown away already.”
That is not a finance failure. That is an agent shipped without a budget primitive.
An alert is not a control
This is where most teams get it wrong. They add observability, wire up a Slack alert at 80 percent, and call it cost governance. It is not. It is a smoke detector in a building with no sprinklers.
The distinction is everything. Token budget enforcement is the set of runtime controls that terminate or pause an agent when it reaches a defined token or cost threshold. Unlike budget alerts, which notify after spending occurs, enforcement stops the agent: no further LLM calls are placed until a human or policy resumes execution.
The 47,000 dollar lesson makes it concrete. Four agents entered an infinite loop in November 2025. They ran for 11 days. The bill was 47,000 dollars. Nobody noticed until it was over. An alert fired into a void. Nothing stopped the next call.
And the counters most teams rely on are fragile. The cap must be enforced at the infrastructure layer, not in application code, because application-layer cost counters reset on process restart and can be bypassed by exception handling. If your budget lives inside a try/except block, you do not have a budget. You have a suggestion.
Why does this keep happening? Because forecasting is broken by default. 85% of companies miss AI cost forecasts by more than 10%, and nearly 25% underestimate by 50% or more, largely because instrumentation stops at the model API level rather than at the individual agent.
Spend is a first-class design constraint
Treat the budget the way you treat auth or rate limiting: a primitive the agent cannot run without. Three controls, set before production, not after the postmortem.
First, a per task token ceiling and a hard step limit. Non negotiable. Second, a cost circuit breaker in the request path that evaluates spend rate before each call. Third, a safe mode fallback instead of a hard crash: degrade to safe mode, narrow execution to read-only tools, no external writes, no further delegation, capped retries, or restricted model paths. Contain the run; do not just kill it.
Now run the Cost of Doing Nothing math, because CODN is the whole business case here. On one side, an afternoon of engineering to enforce a ceiling. On the other, a loop that empties an annual budget by April, or a fleet storm that burns hundreds of dollars in minutes while your dashboard catches up. The asymmetry is not close.
The market already smells this. Gartner predicts over 40% of agentic AI projects will be canceled by 2027 due to escalating costs and inadequate controls. The teams that survive that cull will not be the ones with the cleverest prompts. They will be the ones whose agents cannot spend a dollar they were not given.
Give your agent a budget it cannot exceed. Then give it a prompt. In that order.