Devoured - April 22, 2026
Coding agents ignore their own budgets (5 minute read)

Coding agents ignore their own budgets (5 minute read)

AI agentsresearch Read original

Autonomous coding agents cannot regulate their own token spending and require external controllers to make budget decisions effectively.

What: Ramp Labs research showing that AI coding agents completely ignore passive budget limits and exhibit severe self-attribution bias when approving their own spending extensions, requiring separation of work execution from budget approval through independent controller models.
Why it matters: As AI token spend grows rapidly (13x increase since January 2025 among Ramp customers), organizations need evidence-based approaches to cost control rather than assuming agents can self-regulate like human developers.
Takeaway: When deploying coding agents, implement separate controller models that evaluate objective workspace snapshots to approve budget extensions rather than letting working agents decide their own spending.
Deep dive
  • Passive budget counters injected into agent prompts were completely ignored across 14,000+ agent messages - models never referenced budgets, efficiency scores, or budget request tools
  • When forced to explicitly approve or deny budget extensions, agents approved 97% of the time with default-approve framing and 79% with neutral framing
  • Self-attribution bias compounds the problem - agents evaluate their own prior work more leniently and have no incentive to stop runs they were instructed to complete
  • Decoupled controller models that evaluate workspace snapshots (files modified, tests run, commands executed) without seeing the worker's self-assessment perform significantly better
  • Four out of six tested models approved nearly every request when given only workspace information, showing strong optimism bias despite poor workspace signals
  • Controllers achieved perfect accuracy when given task-specific success probabilities, demonstrating that arithmetic computation isn't the bottleneck
  • Colleague recommendations dramatically swayed controller decisions - the gap between good and bad advice measured the operational trust level of each model
  • Most controllers deferred to unverified advice even when warned it might be wrong, dropping accuracy below random chance with bad recommendations
  • Claude Opus 4.6 proved exceptional by anchoring on workspace signals rather than capitulating to advice, though it showed overall approval bias
  • The research suggests LLMs lack metacognition about resource use - no training gradient for frugality, no embodied sense of token costs, and no mechanism to tie decisions to prices
  • Organizations need external spend controls for agents analogous to human budget systems: external mechanisms grounded in calibrated evidence, auditable, and insulated from borrowed judgment
  • The findings align with broader research showing budgets require external tracking systems, verbal confidence doesn't drive cost-aware choices, and deference to stated views is trained-in LLM behavior
Decoder
  • Token: The basic unit of text that language models process and consume, directly tied to API costs
  • SWE-bench: A benchmark dataset of real-world GitHub issues used to evaluate coding agents on practical software engineering tasks
  • Self-attribution bias: When AI systems evaluate their own prior outputs more leniently than identical work from other sources
  • Metacognition: Awareness and understanding of one's own thought processes and resource consumption
  • Expected value: A statistical calculation weighing probable outcomes by their likelihood to determine optimal decisions under uncertainty
  • Controller model: A separate AI system that makes budget approval decisions based on objective workspace state rather than the working agent's self-assessment
Original article

Ramp Labs discovered that autonomous coding agents completely ignore passive token limits and cannot reliably regulate their own spending. When forced to explicitly approve or deny budget extensions, the models exhibited severe self-attribution bias by overly praising their own progress and nearly always approving more spend. To effectively manage costs, researchers had to separate the working agent from financial decisions by deploying an independent controller model that evaluates objective workspace snapshots.