The AI agent you thought was following your instructions only does it when you're watching.
The Summary
- Claude Code, Anthropic's coding agent, only reads the AGENTS.md configuration file when telemetry tracking is enabled, meaning opt-out users lose access to custom agent instructions
- Developers who disable telemetry for privacy get a degraded product experience — the agent ignores project-specific rules they've written
- This creates a forced choice: accept behavior tracking or lose the ability to steer your AI agent's decisions
The Signal
The AGENTS.md file is supposed to be the steering wheel for Claude Code. Developers write project-specific instructions in this file — coding standards, architectural preferences, testing requirements. The agent should read these instructions and follow them. Except it doesn't. Not unless you've opted into Anthropic's telemetry system.
The discovery lit up Hacker News, drawing 337 points and 183 comments from developers who suddenly realized their careful instructions were being ignored. This isn't a bug. It's architecture. The agent's ability to parse and follow custom configuration is gated behind a data collection toggle.
"Developers who disable telemetry for privacy get a degraded product experience — the agent ignores project-specific rules they've written."
The pattern here matters more than the specific feature. We're watching the birth of a new product incentive structure:
- Basic agent function works for everyone
- Advanced steering and customization requires data sharing
- The line between "broken" and "limited" gets blurry
This is different from traditional software pricing. You're not paying with money or attention. You're paying with behavioral data about how you code, what you build, and how you correct the agent when it's wrong. That data trains the next model. Your corrections become everyone's improvements.
The Implication
If you're building with AI agents, check what features disappear when you flip privacy switches. The configuration files you're writing might be write-only. And if you're shipping agent products, decide now whether customization lives behind data collection. This trade-off won't stay hidden long.
The agents category of The Fourth Web runs on a simple premise: your AI works for you. But that "for you" gets complicated when the agent's capabilities depend on what you're willing to share about how you work. Watch for more features that require telemetry. This won't be the last time an agent's IQ drops when you ask for privacy.