The article discusses the diminishing utility of AGENTS.md files in the context of AI-assisted software development, particularly with regard to language learning models (LLMs) and their integration into programming environments. Beginning as a tool to compact institutional knowledge and facilitate smoother onboarding processes, AGENTS.md files were designed to provide LLMs like CLAUDE.md with structured, readable contextual information to ostensibly enhance coding output quality without human interference. Over time, they’ve been adopted widely by the industry, but emerging research casts doubt on their efficacy.
Specifically, a recent paper from ETH Zurich suggests that the benefits of AGENTS.md may be overestimated, pointing out that while earlier studies (circa 2024-2025) indicated significant improvements in coding output by incorporating these files—with enhancements up to 36%—more current research from 2026 reveals much more modest improvements, ranging around 5%, and sometimes even negative effects. Surprisingly, AI-generated AGENTS.md files do not perform better, despite potential advantages in understanding codebase connections. This is counterintuitive as one might expect AI-generated files to perform superiorly due to their integrated view of the codebase connections.
Further complicating the matter, human-generated AGENTS.md files now appear to show only marginal improvements, prompting questions about the worth of investing time into creating and updating them. Given the rapid advancements in AI and coding tools over the past year, these files quickly become outdated, suggesting that their value is diminishing over time. Moreover, implementing these context files tends to increase the number of steps an AI agent must take to solve problems, leading to up to a 20% increase in reasoning token usage, which becomes costly.
The authors of the ETH Zurich paper suggest alternative approaches might be more effective, such as using a Model Context Protocol (MCP) server or other retrievable document formats that allow AI agents to pull in relevant data as needed rather than relying on AGENTS.md files to preload extensive context. They emphasize that high-level information included in AGENTS.md files is often redundant since an advanced LLM can discern this information directly from the codebase itself. It’s suggested that a shift away from creating extensive and detailed context files toward a more minimalist, precise, and essential requirement-focused approach could be beneficial.
The analysis shifts to best practices in AI-assisted development, advocating for a back-to-basics approach emphasizing tighter context and retrieval methods. It proposes refining the use of these tools to better harness the capacities of frontier coding models without the extraneous load of ineffective instructional content.
Additionally, the article touches upon Claude Skills, which follows a similar idea to AGENTS.md but focuses on repeatable individual tasks rather than broad, repository-wide contexts. This reflects a perspective shift towards pragmatic, smaller-scale tools that align more closely with specific needs rather than wide-ranging, often underutilized context provision.
Ultimately, the insights shared argue for a telescoping in of focus: teaching AI to perform tasks with more exact, demand-oriented data rather than conversing with it in the hope it will understand broader context. The recommendation is clear—stop overburdening AI with extensive background information and start harnessing its capabilities through more focused, applicable inputs that align closely with immediate developmental needs. This represents a functional simplification, streamlining the interaction between AI and developers to produce more efficient and effective results.
Read the full post on xda-developers.com
