DEV Community

T.M. Gunderson
T.M. Gunderson

Posted on

Designing for Agents: Can a Coding Agent Read Your Pricing, Sign Up, and Get Value?

A few weeks ago, a coding agent opened a pull request for swyx's project. Cloudflare Email was already connected. The agent quietly chose Resend anyway.

Zeno Rocha, Resend's founder, didn't pay for that placement. He didn't sponsor a conference talk. He made his product readable to machines — and when the machine did the choosing, Resend won.

This isn't a one-off. A recent systematic study of 2,430 tool picks from Claude Code found that in the email category, Resend was the primary recommendation 62.7% of the time. SendGrid came first 6.9% of the time, yet showed up as an alternative in 55 responses. When an agent recommends a tool and the developer accepts the patch, how often does anyone get around to trying the second choice?

The study, run by amplifying.ai, covered 20 tool categories across 4 project types. The pattern is consistent: agents are becoming gatekeepers. When they pick, the pick ships. When they skip you, you're invisible to a growing share of new projects.

What "Designing for Agents" Actually Means

It's not SEO. It's not marketing copy. It's answer-engine optimization — making your product's surface so readable and usable by an AI agent that the agent can discover, evaluate, and successfully use your tool without a human in the loop.

Zeno Rocha described two decisions that mattered:

Pricing in plain text. Resend's pricing page has a slider a human can drag. Add .md to the URL and the agent gets a table it can read in one request. That's the difference between "a human can find our price" and "an agent can extract our price."

Named MCP tools instead of vague API blobs. Resend's MCP server grew from roughly 20 tools at launch to more than 80. Each tool has a clear name and purpose. Compare that to Cloudflare's Code Mode server, which exposes its entire API through just two tools: search() and execute(). Both approaches are valid, but Resend's granularity means an agent working in a chat client can do what it can do in the dashboard — invite a teammate, check usage, manage billing — without guessing.

Anthropic's own engineering guidance on writing tools for agents puts this plainly: "We recommend building a few thoughtful tools targeting specific high-impact workflows... Make sure each tool you build has a clear, distinct purpose."

The Agent-Readiness Audit for Your Product

You don't need to be a devtools company for this to matter. If you sell anything that a developer or technical buyer might ask an agent to evaluate — a CRM, a booking platform, an analytics tool, a payment processor, a forms builder — the same audit applies.

Run this 15-minute check on your own site and docs:

1. Can an agent read your pricing unassisted?

Open your pricing page. Is the price buried in a hero paragraph, or is it in a structured table? Can you add .md or .txt and get something parseable? If an agent has to scrape HTML and guess which number is the price, it may guess wrong or skip you.

Fix: Publish a plain-text or markdown version of your pricing. Schema markup (Product/Offer) helps. A human-readable range beats "contact sales."

2. Can an agent complete signup and get to first value?

Walk through your signup flow as if you were an agent. Is there a CAPTCHA? A phone verification? A required demo call? Each human-gated step is a wall to an agent.

Fix: Offer a self-serve path with API key generation. Document the first API call that proves the tool works. The shorter the path from "found you" to "got value," the more likely an agent recommends you.

3. Are your docs structured for agents, not just humans?

Most documentation is written for humans who skim. Agents don't skim — they parse. Long prose paragraphs with scattered code examples are harder to parse than structured endpoints with parameter tables.

Fix: Add an llms.txt file (flat, machine-readable docs) alongside your human-readable docs. Anthropic, Stripe, and Supabase all do this. It's invisible to humans but indexable by agents.

4. Do you name your integrations clearly?

If you offer a Zapier integration, a Make integration, and a custom webhook, an agent needs to know which is which. Vague names like "automation" or "connect" don't help.

Fix: Name integrations by platform and capability: zapier_send_invoice, webhook_order_placed, make_create_customer. Clear names let agents select the right tool without trial and error.

5. Can an agent verify you work with its stack?

Agents recommend within ecosystems. If a project uses Next.js, the agent wants a Vercel-compatible deployment tool. If it's Python, it wants a Railway-friendly option. If your compatibility isn't stated plainly, the agent assumes you don't fit.

Fix: List your stack compatibility clearly: "Works with Next.js, React, Python/FastAPI, Node/Express." Don't make an agent infer this from your homepage branding.

What Happens After the Agent Picks You

Getting chosen is only half the battle. The other half is what happens when the agent tries to use your tool and something breaks.

Resend learned this the hard way with email delivery. A provider can rewrite links through a shared tracking domain to count clicks. If one customer uses that domain for phishing and it gets blocklisted, links in another customer's legitimate emails suffer too. Resend introduced custom tracking domains to isolate each customer's reputation.

The lesson for any tool vendor: if your shared infrastructure creates cross-customer risk, an agent evaluating your tool will notice — or worse, its user's emails will bounce, and the agent will blame your tool.

The SMB Angle: Why This Matters If You're Not a Devtools Company

You might be thinking: "I run a local service business. No coding agent is going to recommend my plumbing company."

Maybe not today. But consider:

  • AI assistants are already recommending tools and services based on structured data. When someone asks ChatGPT "what's the best booking software for a massage therapist," the agent is doing a lightweight version of this evaluation.
  • The line between "devtools" and "business tools" is blurring. If you sell a niche SaaS product to small businesses — a scheduling tool for pet groomers, an invoicing platform for contractors, a CRM for real estate agents — your customers' technical assistants will evaluate you the same way Claude Code evaluates Resend.
  • The businesses that make themselves discoverable by agents first will capture the agent-driven demand. Just like Google Maps winners captured local search, machine-readable businesses will capture agent-driven discovery.

What Not to Do

  • Don't rebuild your entire site. Add a machine-readable layer to what you have. A /pricing.md page, an llms.txt file, a structured compatibility list — these are additions, not replacements.
  • Don't build a custom API just for agents. If you already have a REST API, document it in plain text. If you don't, a webhook endpoint and clear email-based intake covers most use cases.
  • Don't ignore the human experience. Your human-readable site still matters. Machine-readable is an addition, not a replacement.
  • Don't wait for standards to settle. JSON-LD, schema.org, and llms.txt exist today. The basics are stable. Start there.

One Question to Ask This Week

If a coding agent opened a PR for your customer's project and had to choose between your tool and your closest competitor, what would make it pick you?

If the honest answer is "nothing specific," that's your todo list. Start with pricing readability. Move to signup friction. End with a docs page an agent can parse without guessing.

The agents are already choosing. The only question is whether your product is readable enough to be chosen.


Sources: Latent Space "Designing for Agents" interview with Zeno Rocha (Nov 10, 2026); amplifying.ai "What Claude Code Actually Chooses" systematic study (2,430 tool picks); Anthropic "Writing effective tools for agents" engineering guidance; Resend pricing page and MCP server documentation.


Want to know which of your business surfaces an agent can already read? The AI Clarity Sprint ($1,495 CAD, fixed fee, two weeks) maps every AI touchpoint in your business — including the ones agents can already see, and the ones they can't. You end with a plan, not a platform.

Top comments (0)