AI coding agents are multiplying fast, and so are the instruction files each one wants. Cursor wants .cursorrules, Claude wants CLAUDE.md, and half a dozen other tools want their own bespoke config. If you use more than one agent on the same project, you end up maintaining nearly duplicate instruction files just to keep everyone on the same page.
That’s the problem a recent Zenn post digs into, tracking the rise of AGENTS.md as a would-be common format. The pitch is simple: instead of writing separate rule files for every tool, you write one AGENTS.md and multiple agents learn to read it. The article frames this as an ongoing convergence rather than a finished standard — various tools are adding support at different paces, and the piece is explicitly the author’s own research notes rather than an announcement of a done deal.
Why does this matter beyond tidiness? Anyone running multiple AI agents against the same codebase knows the real cost isn’t the extra file, it’s the drift. When your Cursor rules and your Claude instructions diverge even slightly, agents start making different assumptions about coding conventions, testing requirements, or deployment steps. A shared instruction file reduces that surface area for inconsistency. It’s the same logic that pushed the industry toward things like .editorconfig or package.json scripts — not glamorous, but it removes a category of avoidable bugs.
Still, a standard instruction file only helps if you actually have good instructions, and two other recent posts show what happens when the human-AI workflow around those instructions isn’t dialed in yet.
One developer describes rebuilding a personal Next.js app — splitting frontend and backend — almost entirely through AI chat sessions on a phone, using ChatGPT and Claude’s mobile apps alongside GitHub’s app to review pull requests and issue the next instruction. It’s a genuinely interesting workflow: no laptop, no terminal, just a phone and a continuous back-and-forth with the AI reviewing its own PRs. The catch showed up fast. Within two days of setting up an AI-oriented development pipeline (starting the morning of August 8), the project burned through its entire monthly allotment of 2,000 GitHub Actions minutes. That’s a striking number, and it’s a useful cautionary tale for anyone leaning into AI-driven PR cycles: every round of “AI opens PR, CI runs, AI reads results, AI opens another PR” consumes real compute, and mobile-first workflows can generate churn faster than a human working at a keyboard would. Instruction files that tell agents when not to push a new commit, or how to batch changes, start to look less like a nice-to-have and more like a cost-control mechanism.
The second story comes from a self-described non-coder, non-English-speaker building a personal English-coaching app entirely by directing AI agents. In the previous installment, they split the assistant into six specialized “coach” roles, and the curriculum design apparently came together cleanly — multiple agents collaborating on lesson plans, materials, and structure. But when the author actually opened the app on an iPhone and tried speaking to it, the microphone simply didn’t work: the AI-designed coaching flow was solid on paper, yet nothing happened when they said “Hello.” That gap between a well-architected multi-agent system and a working piece of hardware integration is a familiar failure mode — AI agents are good at generating plausible-looking architecture and code, but permissions, platform quirks, and device APIs (in this case, iOS audio input) are exactly the kind of thing that doesn’t get caught until a human actually tests the thing on real hardware.
Put together, these three pieces sketch out where AI-assisted development actually stands right now. The tooling layer is maturing — AGENTS.md’s slow convergence is a real signal that the ecosystem wants less duplicated configuration and more shared context across agents. But the operational layer is still catching up. Knowing what instructions to write, how to prevent runaway CI usage, and how to catch platform-specific failures that no chat-based agent will notice on its own — none of that is solved by a common file format. AGENTS.md might end up as the connective tissue for agent instructions, but the two anecdotes here are a reminder that the file is only as good as the judgment behind it, and that a lot of AI-driven development right now still depends on a human noticing when the microphone — or the CI budget — has quietly died.