Coding agents waste tokens because AGENTS.md gets wordy, researchers find

Open-source developers keep accidentally turning their agents into token-spenders. In a new preprint titled “Configuration Smells in AGENTS.md Files: Common Mistakes in Configuring Coding Agents,” researchers affiliated with the computer science department of the Federal Institute of Minas Gerais in Brazil report that configuration smells are widespread, with “Lint Leakage” affecting 62 percent of the files they tested.

The headline villain is simple: duplicative, overly detailed agent instructions. Coding agents often rely on configuration files that summarize expected agent behavior, commonly written in Markdown and named either CLAUDE.md for those using Anthropic models or AGENTS.md for pretty much everyone else. These files include text instructions advising the agent about desired behavior and tool use, and they can get wordy enough to harm both output quality and cost. Based on the researchers’ dataset of 100 popular open-source projects containing an AGENTS.md or a CLAUDE.md file, the authors found the top three “smells” were Lint Leakage (62 percent), Context Bloat (42 percent), and Skill Leakage (35 percent). And they even connect the dots to Anthropic’s own guidance: no more than 200 lines of text, because longer files consume model context and may hinder model coherence.

To understand why boards should care, start with what these context files actually do. In an agent workflow, configuration files act like the agent’s “always loaded” memory on how to behave and how to use tools. If that memory is sloppy, the model has to carry the extra baggage every session. That becomes a cost multiplier, and it can also dilute signal. The researchers define Lint Leakage as instructions that repeat rules already enforced by linters, format checkers, and static analysis tools. In plain English: if a tool will automatically catch and correct a style or correctness issue, telling the model to do it again is wasted effort. It still consumes tokens because the guidance is shipped into the context window.

Context Bloat is the next failure mode. It describes the tendency of developers to overspecify code agent behavior. The researchers warn that bloated configuration files increase token consumption, raise costs, and reduce the visibility of important instructions. This is where the earlier Anthropic line limit gets relevant. The authors are not just hand-waving about “too much text.” They point to that recommendation that configuration should not exceed 200 lines because model context is finite. When your agent’s instruction file gets large, you can end up with less coherent priorities, which means the model may still “know” what you wanted but struggle to execute it cleanly.

Then there is Skill Leakage, which is sneakier because it is less about verbosity and more about clutter. Skill Leakage occurs when rarely used tools or practices get added to the AGENTS.md file, which gets loaded in every agent session. The researchers argue these instructions would be better in a separate skills file (for example, SKILLs.md) that gets loaded only when needed. The second-order impact here is easy to miss: adding seldom-used capabilities not only expands context unnecessarily, it can distract agents from other tasks, because the model’s attention and prompt budget are pulled toward information that may not apply.

The study also lists other “agentic odors” beyond the top three. Blind References happen when configuration files reference external documents (for example, via URLs) without explaining when that resource becomes relevant. Init Fossilization refers to configuration details set up upon a project’s initialization that are no longer relevant later. Conflicting Instructions occur when agent directives contradict each other. Even if your team is not writing code smells, these problems can creep in through iterative documentation habits, copy-paste, and “we’ll probably need this someday” thinking.

One of the most executive-friendly findings is how common this is. The authors say they found at least one of these six smells in 91 of the 100 AGENTS.md files tested. They conclude that developers could benefit from catalogs and tools designed to spot configuration issues in agent configuration files, using their preprint paper, “Configuration Smells in AGENTS.md Files: Common Mistakes in Configuring Coding Agents,” as the foundation for that argument.

And this is not happening in a vacuum. The Register also points to earlier work from ETH Zurich on context files for agents. That earlier examination found that developer-generated instructions raised costs and only improved code performance about 4 percent, while LLM-generated instructions had a small 3 percent negative impact on agent-generated code. The ETH Zurich authors concluded that “unnecessary requirements from context files make tasks harder,” and that human-written context files should describe only minimal requirements. Put together with the new findings, the pattern is hard to ignore: more instruction text is not automatically better. It can actively make agent execution more expensive and less reliable.

For decision-makers, the operational takeaway is straightforward but uncomfortable: agent configuration hygiene is now part of cost control, quality management, and risk reduction. Token waste is a margin issue. Distracted agents are a performance issue. Conflicting or stale instructions are an execution risk. And as agents become embedded in software delivery, these “smells” can quietly erode the return on every model call. The strategic stakes for executives and boards are simple: if your product roadmap depends on coding agents, prompt and config discipline is not an engineering nicety, it is a lever that determines whether your automation pays for itself.