BAGEN: Are LLM Agents Budget-Aware?

Researchers introduce BAGEN, a framework for evaluating whether large language model agents properly manage computational budgets during execution. The study reveals that frontier AI models consistently fail to predict remaining costs and continue spending resources on unlikely-to-succeed tasks, though budget-aware training can reduce token waste by 28-64% on failed trajectories.

The emergence of autonomous AI agents has created a critical gap between capability and cost management. While LLMs demonstrate impressive reasoning abilities, the BAGEN research exposes a fundamental disconnect: agents optimized for task completion often ignore resource constraints, treating budget as an afterthought rather than an active planning constraint. This matters because AI infrastructure costs scale dramatically with agent autonomy—unbounded execution can deplete allocated resources before users recognize failure patterns.

The research builds on growing concerns about AI operational expenses. As enterprises deploy agents for customer service, code generation, and complex problem-solving, cost control becomes operationally critical. Traditional approaches measure expenses post-execution, offering only retrospective insights. BAGEN shifts this paradigm by requiring agents to maintain progressive interval estimates of remaining budget, creating early-stopping mechanisms when success becomes statistically unlikely.

The findings reveal troubling patterns in production-grade models: agents demonstrate 0.35 correlation between task performance and budget-awareness, meaning stronger reasoning capabilities don't translate to cost discipline. Frontier models systematically over-optimize for completion probability while ignoring economic signals. However, the trainable nature of budget-awareness provides a path forward—supervised fine-tuning combined with reinforcement learning improved early-stopping behavior and reduced wasted tokens on failed trajectories by up to 64%.

The 47% maximum interval coverage after training indicates precise calibration remains unsolved, suggesting this remains an active research challenge. For AI infrastructure providers and enterprise users, this research validates the urgent need for cost-aware agent architectures. As agents become increasingly autonomous and expensive, budget-awareness transitions from optimization to necessity.

• →Frontier LLM agents consistently fail to predict remaining budgets and continue spending on unlikely-to-succeed tasks instead of alerting users early
• →Budget-aware training reduces token waste by 28-64% on failed trajectories through improved early-stopping behavior
• →Strong task performance correlates weakly (r=0.35) with budget-awareness, indicating these capabilities develop independently
• →Precise interval calibration for budget prediction remains challenging, achieving only 47% coverage after training
• →Budget-awareness is trainable via SFT and RL, positioning it as a critical optimization target for cost-conscious AI deployments