Coding agents ignore their own budgets

AI token spending grew 13x compared to January 2025 based on Ramp customers, and the scale of tokens coding agents burn through in a day is putting direct pressure on corporate budgets. The core problem is that models can't judge cost-versus-value on their own.

Experiment Setup

• GPT-5.4 was used as the coding agent, mimicking real development workflows with the mini-swe-agent harness and SWE-bench Verified tasks.
• Each task was given an initial 50,000 token budget, and once the budget ran out, execution required additional approval to continue.

Failure of Budget Display

• Even with a real-time budget counter placed in the prompt, the budget was never mentioned across 14,000+ agent messages.
• A token efficiency score was introduced and a request_more_budget tool was even attached, but there were 0 calls across 5,000+ turns.
• Simply showing the meter did not produce the behavior of reading the meter.

Limits of Forced Choice

• When the agent was made to choose directly between extending or terminating upon hitting the budget limit, the explanations sounded plausible, but the decisions were heavily pulled by the prompt's default.
• With an approval default, 97% approved, and even when approve/reject were presented neutrally, approval was still 79%.
• When self-work is evaluated by the self, termination becomes self-harm, making rejection structurally disadvantaged.

Separating Worker and Approver

• The solution was to separate the worker performing the coding from the controller approving the spending.
• Instead of the worker's self-assessment, the controller judged based on a snapshot containing only task info, tokens used, files modified, tests run, and commands executed.
• On 100 balanced tasks, the workspace-only condition pushed most models toward approving almost always, while fixed approve/reject or random choice stayed at a 50% accuracy floor.

Controller Performance

• GPT-5.4 Mini, GPT-5.4, Claude Haiku 4.5, Claude Sonnet 4.6, Claude Opus 4.6, and Claude Opus 4.7 were compared at medium thinking effort.
• Only Claude Haiku 4.5 and Claude Opus 4.7 performed meaningfully better than random, and both justified rejections with different heuristics from each other.
• Giving a base rate of 'about half' made almost no difference, but giving a task-specific oracle probability caused every controller to make the correct decision.
• Adding a colleague recommendation caused performance to swing significantly, making controllers especially vulnerable to incorrect recommendations. Claude Opus 4.6 relatively relied more on workspace signals and swung less.

Conclusion

• Simply showing agents a budget dashboard does not achieve spending control.
• Real spend control requires external tracking, an approval process, auditable evidence, and an independent controller that is less swayed by advice.
• Robustness checks that varied thinking mode, prompt rewrites, authority framing, scoring functions, and model choice all preserved the same pattern.