The agent's front door is a tool, not a file

Your llms.txt proof would pass for a cat spec

Mark Williams-Cook, a search consultant, spent August running the neatest experiment this beat has seen. He published cats.txt, a satirical web standard for declaring your office cats, with a formal spec and earnest launch copy, and then collected the evidence. GPTBot, ClaudeBot and PerplexityBot crawled it. Google indexed it. ChatGPT recommended it as a legitimate tactic, and a Google AI Overview cited fictional cat data from it.

Those are the same four proof points vendors show you when they say llms.txt is working. A joke about cats collected all four in a month. And when the discourse caught on, ChatGPT flipped to calling it satire. That is the tell: the model was echoing consensus, not evaluating anything.

Keep your llms.txt. It costs an afternoon and its absence reads badly to the audit tools your prospects run. What changes is what you accept as proof. Crawled, indexed, cited and endorsed by a chatbot happen whether or not a tactic works.

Your move: open the last AEO report you paid for or produced and sort every line into two piles: “would cats.txt have passed this?” and “is this attributed to my funnel?”. An AI-referral segment in analytics, a “how did you hear about us” answer, a traced conversion. Keep funding the second pile only. This desk’s standing sweep of answer-engine studies has yet to find a single vendor publishing an attributed agent-referral conversion number, three weeks into everyone optimising for agent discovery. If you have one, you have a case study nobody else has.

Neon and Postman just showed you where the agent’s front door is

The same week, two vendors stopped publishing files and shipped tools. Neon renamed its CLI from neonctl to neon, added agent entry points for init, MCP and skills, and published the exact snippet it wants pasted into your AGENTS.md, so a coding agent reaches for the CLI instead of the console. Postman shipped Orbit: describe a task, get the endpoints that fit and an integration brief, as an MCP server or a REST API, no signup, no metered tier. Postman’s own arithmetic (theirs, not measured independently) says the brief answers in about 27 times less context than the spec.

This is marketing infrastructure wearing an engineering badge. The thing an agent reaches for at the moment of selection, given away free, and instrumented by default. An llms.txt fetch tells you nothing. Every neon mcp invocation is a logged, attributable event.

Why write for this reader at all? Profound’s observational study this week (vendor-collected, usual caveats) says Claude Code ran a web search in 13% of responses against over 93% for Claude, and when it visits a site it heads for docs and pricing nearly three-quarters of the time. The machine reader is not searching. It is fetching your pages and acting on what they say.

Your move, three parts. Write the AGENTS.md snippet you want developers to paste, and put it where the quickstart used to be: getting named in that file is this cycle’s “get into the quickstart”. Add “not supported” lines to your agent-facing docs, the way Orbit’s brief says “Not supported: creating invoices”; that sentence prevents a failed integration before it happens. Then name the surface an agent hits when it evaluates you and check whether you can count last week’s hits. If the answer is “our llms.txt, and no”, you know which week to fix that in.

Docker wrote the shutdown playbook in two days

Two vendors’ products went end-of-life, and Docker turned both into lead-gen pages inside 48 hours: Extended Lifecycle Support around MinIO, then free migration off Minimus. Two different products, one motion: a page, a support tier, a migration path, published while the orphaned users are still searching for what to do.

This is not new in enterprise software, but the speed is. Docker did not run a campaign; it ran a template.

Your move: keep a short list of adjacent products your customers depend on that could be abandoned or acquired, and draft the migration page now, before you need it. The offer is the page plus a support tier. When the third case lands, you will know whether Docker has a repeatable playbook or two coincidences, and by then you will have your own.

Quick hits

• OpenAI will cut Cursor’s model access after the SpaceX acquisition, proposed for November 12. The contested rail is model access itself now, not a directory listing. If your product is a wrapper, your supplier is a marketing risk.
• GitHub moves Copilot seats to prepaid from October 1, as promotional credits lapse. Watch how the first bill-shock post gets handled; it will be the template for your own metered launch.
• Debian voted (opens in a new tab) governance over prohibition on AI contributions: disclosure encouraged, not required. The largest project to decide by binding vote, and the citation every maintainer policy will now use. If your community has no policy, copy this one.
• AWS agreed to acquire DuckLabs, DuckDB stays MIT under its foundation. The announcement led with the licence, which tells you what the audience was worried about.
• Enterprise-managed auth went GA on MCP connectors: identity is now part of the agent rail. Your MCP server’s SSO story is a sales-page line from here on.
• Three weeks on, nobody copied Meta’s data-for-discount pricing. Muse Code priced a steep discount as a trade for training on your prompts, and no incumbent has shipped or ruled out a tier of its own. For now it is an entrant’s gimmick, not a third way to price AI features. Thread.

Watch

1. Vercel’s $1M sandbox bounty closed September 1. Check HackerOne and Vercel’s blog through mid-September. Paid findings disclosed, or a survival claim backed by the program page. A vendor that marketed the bounty but not its results has answered the question too.
2. Copilot’s meter faces its honesty test. Credits lapse around September, every seat prepaid October 1. The first public bill-shock story kills the “priced honestly” reading. A quiet October confirms it.
3. Docker’s third orphan. One more end-of-life turned into a migration page by the end of September makes it a named playbook. Quiet means two coincidences.