Developer Relations Docs Have a New Primary Reader

← Back to Blog

Fern, a developer docs platform, claims a 90%+ reduction in token usage. It serves clean markdown to LLM bot traffic instead of full HTML pages. Bots already sent more traffic than humans, so Fern made the change. Mintlify now auto-hosts an MCP server for every docs site. GitBook added llms.txt support in January 2025.

These are infrastructure decisions, and infrastructure follows usage. Someone measured the traffic and decided that AI agents reading documentation were worth the investment.

Developer relations teams must now decide what to do about it.

The gap between what agents need and what DevRel docs provide

The Stack Overflow Developer Survey 2025 found that 66% of developers cite “AI solutions that are almost right, but not quite” as their top AI pain point. A separate 2025 industry analysis found that 65% of developers say their AI coding assistant misses relevant context during code review and refactoring.

Your docs cause that gap. AI assistants fetch whatever docs exist. If your docs are stale, the agent gives a wrong answer. If your docs are written for a reader who skims, the agent still gives a wrong answer.

Developer advocates and technical writers write for a human reader. A human reader can ask a follow-up question and forgive some ambiguity. Agents cannot do either. They take only what fits in the context window and generate an answer from it.

Thin or outdated docs produce a confidently wrong agent answer.

Context rot is not a model problem

AI research calls this issue context rot. A 2025 study from Chroma tested 18 frontier models. Every model showed performance loss as input length increased. Context rot starts well before the context window fills up. It happens on simple tasks. It happens across all major models.

A related effect causes accuracy drops of 30%+ for information placed in the middle of a long context. Agents do not read top to bottom, and they do not weight every part of the input equally. They lose track of information buried in the middle.

Length works against developer relations docs. Every stale endpoint in your docs lowers the quality of agent answers. Every deprecated parameter in your docs does the same.

More is not better

When agents start reading your docs, the natural instinct is to write more. Teams add more guides. Teams add more context files. A February 2026 paper from ETH Zurich shows this instinct is wrong.

The ETH Zurich study examined AGENTS.md files across more than 60,000 open-source repositories. AGENTS.md is a context file format. It gives AI coding agents product-specific instructions. The study found that these context files did not generally improve task success rates. Adding them increased inference cost by more than 20% on average.

The researchers recommended a simpler approach. Omit LLM-generated context files. Include only specific instructions the agent cannot infer on its own, such as which test runner to use and which commands to avoid. Add any constraints that are unique to the project.

The pattern holds in practice, too. More content raises the chance that an agent’s context window fills with irrelevant material. Signal-to-noise ratio matters more than completeness.

What DevRel teams should change

DevRelCon 2026 opened a dedicated track for writing docs for both human and agent readers. This new framing matches the research. Docs need a higher standard of precision and a lower tolerance for drift.

Three changes have the most impact.

Task-scoped content over complete reference. Agents retrieve docs for one specific task. A guide that answers “how do I paginate the responses API” beats a reference page that covers the entire responses object with every parameter. Break long pages into task-scoped units. Give each unit a specific title that matches how developers phrase the question. The practical guide to optimizing docs for agents explains how agents access your content and which structural changes matter most.

Freshness markers. Agents cannot tell when your docs were last updated. If your docs describe an API that changed six months ago, the agent gives your user instructions for the old behavior. Explicit version tags and last-updated dates give agents a signal about reliability. These markers also create internal pressure to keep content current.

Minimal, high-signal context files. If you maintain a CLAUDE.md, AGENTS.md, or llms.txt for your product, keep it short. The ETH Zurich data suggests under 200 lines as a reasonable ceiling. Focus on what is non-obvious. Cover authentication quirks and rate limits that are lower than developers expect. Do not restate what the reference docs already say.

Stale docs now fail two audiences

Documentation decay has always been a problem for developer relations teams. Products change faster than documentation workflows can track, so docs go out of date. In the past, this produced one frustrated developer who found the wrong answer and tried something else.

Now stale docs fail two audiences. A developer using an AI coding assistant to integrate your API gets a wrong answer from the assistant. The developer may not know the answer came from your docs. The developer sees a confident response and tries to implement it.

Zylos AI’s 2025 analysis attributed 65% of enterprise AI failures to context drift or memory loss during multi-step reasoning. Agents working from bad inputs fail more often than agents working from accurate inputs.

The docs your DevRel team publishes are now part of the reasoning chain for millions of developer tasks per day. Keeping docs accurate and current has always been the goal. The stakes are now higher. Documentation drift is harder to catch than it looks.

If you want to see where your docs stand today on accuracy and coverage, Promptless monitors for documentation drift. Promptless flags gaps and outdated content before they reach developers or the agents working on their behalf.