Token Budgets: An Empirical Catalog of 63 LLM-Agent Budget-Overrun Incidents, with an Affine-Typed Rust Mitigation as a Case Study

2026-06-04 · Source: cs.SE updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

Sajjad Khan's research presents an empirical catalog of 63 confirmed production incidents of LLM-agent budget overruns, drawn from 21 orchestration frameworks between 2023 and 2026. These incidents, backed by GitHub issues and documented dollar losses, are categorized into an eight-cluster failure taxonomy, with a two-human inter-rater reliability of Cohen's κ=0.837 on the full 113-entry sample. As a mitigation, the author introduces "token-budgets," a 1,180-line Rust crate that uses affine ownership to transform common budget integrity violations like cloning, double-spending, or use-after-delegation into compile errors. Evaluation across five production runtimes and three providers demonstrated zero cap violations and zero false refusals. The crate's core value lies in ensuring non-bypassability under operator error in multi-agent delegation, rejecting the M-delegation-fanout race at compile time, unlike runtime alternatives.

Key takeaway

For AI Engineers developing new Rust-based LLM agents, integrating the "token-budgets" crate is crucial for preventing costly budget overruns. This approach ensures compile-time integrity, making common errors like double-spending or mis-delegation impossible, thereby providing a non-bypassable session-cumulative cost cap. You should consider this discipline for multi-provider or multi-agent deployments where runtime checks alone are insufficient against operator error, enhancing capital efficiency and operational safety.

Key insights

LLM-agent budget overruns are a recurring production failure, mitigated by compile-time affine typing for non-bypassable cost integrity.

Principles

• Compile-time integrity prevents budget misuse.
• Runtime caps bound dollar consequences.
• Affine ownership ensures non-bypassable bookkeeping.

Method

The "token-budgets" Rust crate operationalizes affine ownership via a Budget API. It uses self-consuming methods (spend, split, merge) to make aliasing, double-spending, and use-after-delegation compile errors.

In practice

• Use Budget::new with a capability token.
• Employ budget.spend for cost debiting.
• Utilize budget.split for sub-budget delegation.

Topics

• LLM Agents
• Budget Management
• Rust Programming
• Affine Types
• Compile-time Safety
• Cost Overruns
• Resource Management

Code references

• sajjadanwar0/token-budgets

Best for: AI Architect, NLP Engineer, CTO, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• WorkstreamBench: Evaluating LLM Agents on End-to-End Spreadsheet Tasks in Finance — WorkstreamBench evaluates LLM agents on end-to-end financial spreadsheet tasks, revealing current limitations in professional-grade output.
• The Pulse: AI token spending out of control – what’s next? — Exploding AI agent token spend and vendor demand outstrips budgets, while Meta faces internal morale issues.
• Why Rust is the Ideal Language for Vibe-Coding — Daniel Szoke, Sentry — Rust's strict compiler constraints, though challenging for LLMs initially, enhance production code safety by preventing common errors.
• Reasoning Budgets vs. Structured CoT: Controlling LLM Thinking Tokens — Controlling LLM reasoning traces via budget or grammar can reduce token cost but risks accuracy.
• Further Notes on Our Recent Research on AI Delegation and Long-Horizon Reliability — LLMs can accumulate semantic degradation in long, delegated workflows, despite strong short-horizon performance.
• The Invisible Failure Mode of Agentic AI — Agentic AI failures are often invisible resource issues, not reasoning flaws, requiring real-time, local observability.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.SE updates on arXiv.org.