A New Tool Just Graded Whether AI Agents Can Use Your Product. Should Your YC W27 Application Treat Agents as a Customer?

This morning a tool called AX Check landed on the Hacker News front page with a simple premise: it sends three coding agents to your website and grades how far they get trying to onboard to your product, end to end, with no human help. Snyk, Increase, and Loops all score a perfect 100. Mercury, a company that banks half of YC, scores 60 because its docs index is broken and an agent can't confirm pricing.

Read that last sentence again. A benchmark now exists where a fintech serving thousands of startups loses to a transactional email tool, because an AI agent got confused.

If you're applying to YC Winter 2027, this is not a curiosity. It's a preview of a question that is going to show up, in some form, in your application review and your interview: can AI agents use your product, and does your growth model assume they will?

What is agent experience (AX), exactly?

Agent experience is the emerging discipline of making your product usable by AI agents the way UX made it usable by humans. Concretely, AX Check (built by Gauge, an agent-infrastructure company) grades things like:

• Whether your docs are machine-readable and indexed (llms.txt, clean markdown, no login walls)
• Whether pricing and setup steps are discoverable without a sales call
• Whether there's a CLI, API, MCP server, or Skills package an agent can actually drive
• Whether three real agent sessions can sign up, configure, and reach a working outcome

The leaderboard is telling. The top scorers are developer-first companies like Kernel, Loops, Increase, and Snyk. The companies agents struggle with are the ones whose funnels were designed for a human with a credit card and patience: gated docs, "book a demo" pricing, onboarding flows that require a phone number and a marketing quiz.

Why would YC care whether agents can use your product?

Three reasons, all grounded in what's already happened this year.

First, agents are becoming a real acquisition channel. When a founder asks Claude or ChatGPT "set me up with error monitoring," the model picks a tool, reads its docs, and completes signup. If the agent stalls on your site, it picks your competitor. This is the same logic as GEO (generative engine optimization), but one level deeper: it's not enough for AI to recommend you, AI has to be able to finish the job with you. We covered being "in the weights" back in June; AX is what happens after the model remembers you exist.

Second, YC's own portfolio is telegraphing the shift. This week's HN front page had Cekura (YC F24) hiring forward-deployed engineers for AI agent testing, and Cactus (YC S25) launching Needle 3, tiny 8-29MB automation models built to drive software directly. Yesterday we wrote about Skillsync (YC W26) and the agent-skills economy. The pattern across batches is consistent: YC keeps funding companies that assume software's next billion users are agents.

Third, it changes the "how will you grow?" answer. The YC application asks how users will find you. In 2024 the strong answer was SEO or a bottoms-up developer motion. In late 2026, a differentiated answer is: "Agents can onboard to our product unassisted; here's our AX score and our llms.txt; 15% of our signups already come through agent sessions." Almost nobody can say that yet. That's exactly why it's worth saying if it's true.

Does this apply if you're not a dev tool?

More than you'd think. The first wave of AX is developer products because coding agents are the agents that exist at scale today. But the same mechanics are coming for anything bought online. If you sell HR software, an agent tasked with "find us a PTO tracker under $10/seat" will hit your pricing page. If it can't find a number, you're not in the consideration set, and no human ever knew you were excluded.

The honest caveat: if you're building a consumer social app or a hardware company, AX is not your W27 story, and bolting it on will read as trend-chasing. YC reviewers have seen every buzzword this cycle. An AX claim only works if agent-driven adoption is plausibly your channel.

What should actually go in your W27 application?

Keep it concrete and verifiable, the way YC likes everything:

In the "how will you get users" question: if developers or businesses buy your product, state plainly whether an agent can complete onboarding today. If you've measured it, give the number. "We score 100 on AX Check and see roughly 1 in 10 signups complete with no human session activity" is one sentence and says more than a paragraph about your content strategy.

In the product description: mention your MCP server, API, or CLI only if it's live. YC partners increasingly test-drive products during review. An agent-usable product can literally be evaluated by the reviewer's own tools. A dead link to docs is worse than no claim at all.

In the competition question: if your incumbent competitor is a Mercury-style company with gated pricing and human-only onboarding, that's a real structural advantage worth naming. Incumbents can't ungate their enterprise funnel without blowing up their sales org. You can be agent-native from day one.

The one-hour version before you submit

You don't need a quarter of AX work to say something true. Run your own site through AX Check, fix the embarrassing findings (publish pricing, add llms.txt, unblock your docs from login), and re-run it. That's an afternoon. Then have someone who's seen a lot of YC applications pressure-test how you're framing it, whether the claim reads as a real channel or as buzzword decoration. Getting a YC alum to review your draft on YC Roaster is the fastest way to find out which one you've written.

The W27 batch will be full of founders pitching products for agents. The quieter opportunity is making sure your product works with them, and being one of the few applications that can prove it.