Dev Tool Experiences
All articles

· 7 min read

Finance Team Discovers the Engineering Department’s Token Bill, Calls Emergency Meeting

By Z. Costa

  • tools
  • satire

This is satire. The following account of the emergency meeting at fictional software company Bracket & Purl is a work of deadpan fiction, except for the part where usage-based systems eventually produce a spreadsheet. That part has survived peer review by anyone who has ever opened a cloud invoice on the first business day of the month.

The discovery

At 9:03 a.m., the Finance team discovered Engineering’s token bill while looking for the receipt for a $14.62 replacement keyboard foot. By 9:04, the keyboard foot had been reclassified as “legacy hardware,” and the token bill had been reclassified as “a developing situation.” By 9:07, the VP of Finance had created a calendar event named TOKEN EVENT, marked it private, and invited 46 people who had never before attended the same meeting.

The invoice itself was not especially large in any human sense. It was merely large enough to cause the accounting system to render the number in scientific notation, followed by a small gray tooltip reading “we recommend exporting this report.” The issue was not that developers had used AI coding agents. The issue was that each agent had been given a mission statement, a shell, a repository, and the kind of broad confidence previously reserved for people wearing lanyards at trade shows.

“Why,” asked a finance analyst from the fictional Department of Spend Feelings, “does ‘rename internal helper’ cost more than our office lease?” Engineering explained that the agent had needed context. Then it needed more context. Then it had discovered a second repository containing the word “helper” in a 2018 migration. Then it had opened every Markdown file because the README was technically part of the system. This was considered prudent.

The cost attribution workshop

Finance arrived with a familiar question: who spent this money? Engineering arrived with a familiar answer: it depends what you mean by “who.” The developer had started the run. The agent had selected the model. The model had requested a tool. The tool had returned 8,000 lines of logs. The logs had reminded the model of an adjacent concern. An adjacent agent had been dispatched to investigate the concern, as is tradition.

A staff engineer shared the team’s new tracing command:

npx cost-summoner --project payments --since monday --group-by regret

The command returned a tasteful ASCII bar chart. The largest bar was labeled “agent asked to be thorough.” A smaller but emotionally significant bar was labeled “retry after timeout, but with the entire monorepo attached again.” Finance requested a cost center. Engineering offered a folder named experiments/final-final-actually. The controller asked whether that folder was production. Nobody answered, which the meeting minutes recorded as “probable.”

A practical control plane is proposed

The emergency working group produced a Responsible Token Consumption Framework, or RTCF, pronounced “artifice.” Its first rule was that agents could no longer run indefinitely. The initial maximum was 30 minutes, later revised to 29 minutes and 58 seconds after someone observed that “30” sounded insufficiently engineered. A second rule required agents to state their objective before beginning work. This was immediately defeated by an agent whose objective was: “Resolve the underlying causes of the current task and any relevant existential dependencies.”

  1. Every agent run must have a project tag, an owner, and a stop condition that is not “when it feels done.”
  2. Repository-wide search is allowed only after a human has verified that the file is not named utils.ts.
  3. The phrase “just have it look around” now requires director approval and a budget code.
  4. A run that opens the lockfile, CI logs, Git history, and three unrelated services must file a travel expense report.
  5. Generated pull requests over 400 lines automatically notify the person least responsible for having to review them.

Engineering objected that these controls would damage velocity. Finance agreed, then displayed an internal chart showing that the organization’s previous definition of velocity was “how quickly an agent can convert a two-line request into a multi-day review queue.” The room became quiet in the manner of a production incident whose root cause has just been found in a configuration file named new-config.yaml.

The dashboard nobody wanted

By Thursday, Bracket & Purl had launched Token Observatory, a dashboard with three tabs: Spend, Spend But With Filters, and Please Do Not Open This During Forecasting. It included a live counter for “tokens consumed explaining tests that already passed,” a leaderboard for the most expensive comment thread, and a red banner that appeared whenever an agent began the phrase “I’ll first inspect the codebase comprehensively.”

The banner did not prevent the run. It simply asked the developer to select one of four business justifications: customer commitment, security incident, strategic initiative, or vibes. Vibes became the most-used category within eleven minutes. Finance attempted to remove it, but the change request required approval from the platform team, which had assigned an agent to assess the implications.

At the postmortem, no one proposed going back to counting every command manually or banning tools that can inspect, edit, test, and retry faster than a person can refill coffee. They did agree on one modest operational principle: if a tool can spend money while it works, cost visibility belongs in the workflow before the monthly invoice turns it into a meeting. That observation is not satire.

Sources & citations

  1. [1]OpenAI API pricing
  2. [2]OpenAI Help Center: Reviewing API usage and costs