# Agents are choosing the stack now. We raised $1.3M to help devtools get picked.

*Manicule raised a $1.3M seed co-led by Y Combinator and Broom Ventures. Why AI agents now pick developer tools, and how we help devtools get recommended.*

By Naman Bansal, CEO and co-founder · November 7, 2026

A year ago, if you wanted to know how developers found your product, you opened Search Console and your referrer logs. Today a growing share of that decision happens somewhere you can’t see: inside a ChatGPT conversation, or in a coding agent’s terminal while it scaffolds a project. The developer asks for “a database for this app” and accepts whatever the agent installs.

We just raised $1.3M, co-led by Y Combinator and Broom Ventures, with angels including Sam Bhagwat, Theo Browne and Kevin Moore, to work on that problem. In this post I’ll explain what we think is going on, why we call the work Agent Relations, and what we’re doing with the money.

## What changed

The clearest evidence comes from founders who watched it happen in their own dashboards.

In February 2025, Paul Copplestone posted Supabase’s weekly signups with the note [“much of the growth from AI Builders/Platforms.”](https://x.com/kiwicopple/status/1894251335224320000) By October, Supabase was launching more than half a million databases a week, [more in one week than in the first 3.5 years of the business](https://x.com/kiwicopple/status/1976411278609699148). In August 2026 he said the number of databases launched by hand was [“converging on zero.”](https://x.com/kiwicopple/status/2084477911172841493)

Guillermo Rauch shared in April 2025 that [ChatGPT refers 10% of new Vercel signups](https://x.com/rauchg/status/1910093634445422639). Han Wang at Mintlify posted that [AI agents and crawlers made up 48% of the docs traffic they measured](https://x.com/handotdev/status/2039734705634058677). And in June, Zeno Rocha posted that Resend’s OpenAI traffic had tripled, and summed it up in a line we’ve repeated ever since: [“agents are choosing the stack now.”](https://x.com/zenorocha/status/2062539292035915973)

Two things are happening at the same time. Assistants like ChatGPT, Claude and Gemini answer the question “what should I use for this?” And coding agents like Claude Code and Cursor go further: they pick a tool, install it, and write the integration. In both cases the agent reads what a developer used to read (your docs, your README, Reddit threads, comparison pages, benchmarks) and narrows the field to one or two names. Most of the time the developer goes with it.

Aaron Epstein, a general partner at Y Combinator, put it this way when we announced the round: “It’s no longer about humans deciding what tools to buy – agents are making these decisions.”

For a devtools company, that’s a distribution channel you didn’t have three years ago. Very few teams are managing it on purpose yet.

## Why we call it Agent Relations

DevRel exists because developers don’t respond to ordinary marketing. They trust docs that work, honest comparisons, and people who answer questions in their communities. Agents lean on exactly the same material. What changed is who reads it first.

So we call the work Agent Relations, or AgentRel: [making your product easy for AI agents to discover, recommend, and use](/agentrel). We didn’t invent the phrase. The New Stack used “From DevRel to AgentRel” as a heading in June 2025, and Quinton Wall and RedMonk’s Stephen O’Grady have written about agent relations since. We’re the first company built entirely around doing the work.

In practice, every recommendation an agent makes passes three tests, and the work splits the same way:

- **Found.** Does the agent come across you at all? Assistants search the web before most answers, so you need content that answers your buyers’ real questions, published where agents read it, including Reddit and LinkedIn.
- **Chosen.** Does it pick you over the rest? That depends on what other sites say about you: the comparison pages and threads assistants already cite, original benchmarks, and correcting things assistants get wrong.
- **Used.** Does the code it writes work? A coding agent follows your docs to the letter, so a broken quickstart becomes a broken integration in someone’s project.

## What’s broken about how teams handle this today

Most companies that notice the shift buy an AI visibility dashboard. I wrote about my frustration with these in [our manifesto](/manifesto): many of them ask a model a question once, print a percentage, and call it your visibility.

We tested how much that matters. A single AI answer misjudged a brand’s real recommendation rate by a median of 14 points. Thirteen answers cut the error to 7, and half of all brand and question pairs needed 18 or more runs before the number settled ([the full method is in our research paper](/research/manicule-prompt-tracking-methodology.pdf)). If you’re looking at one run per question, a lot of what you’re seeing is noise.

The second problem is that a dashboard stops at the number. The work that changes it, rewriting a broken quickstart, publishing the benchmark, getting into the threads assistants cite, still lands on your team.

## Two ways to work with us

That’s why Manicule is two things, and you can use either one on its own.

- **Our team does the work.** We run [Agent Relations as a service](/services): strategy, docs, content, benchmarks and distribution. We own the work end to end, and you review everything before it ships.
- **Our software measures it.** [Agent visibility](/agent-visibility) runs every question many times on every assistant and shows you the answers behind each score, so you can tell a real drop from a bad roll. You can sign up and track your own prompts today.

Most teams start with one and add the other. The software tells you where you’re losing; the team fixes it and the software shows whether the fix worked.

## What we’ve seen work

The engagement that shaped most of our thinking is Supermemory. Shreyans, my co-founder and our CTO, was Supermemory’s founding engineer before Manicule. He built their consumer app and MCP server to more than 100,000 users, so we saw the problem from the inside before we worked on it from the outside.

Our team rebuilt Supermemory’s docs, wrote their content, and distributed it across search and Reddit. The results: more than a million SEO impressions, the top two Google results for “mem0 vs supermemory”, over 200,000 views on our Reddit benchmark posts, and a 30% lift in answer success after the docs rebuild. You can read [the full story here](/case-studies/devrel).

A few things from that work became the way we operate:

- Benchmarks and honest comparisons beat volume. One post with real numbers did more than a dozen generic ones.
- Reddit rewards data and punishes pitches. Our benchmark post did well; a plain link to a case study got downvoted.
- Docs need to be tested the way an agent uses them. We check every snippet against the live API and run coding agents through real tasks before calling a rebuild done.

Since then we’ve worked with teams including Greptile, Rootly and PromptLayer.

## What the money is for

Most of it goes to engineering. That means the tracking software, and the internal agents that handle the repetitive parts of content and docs work, with a person reviewing everything that ships. The rest goes to people who can work directly with our customers’ technical teams. If that sounds like you, [we’re hiring](/careers).

We’re starting with developer tools because that’s where agents already make the most decisions, and it’s where we’ve done the work. Other kinds of software come later.

## What we’ll publish here

This blog is where we’ll share what we learn from doing this work: how often agents actually recommend a given tool, what moves that number, and what doesn’t. Our next post goes deeper on the measurement problem, with the data on how many runs it takes before a visibility score means anything.

If you run marketing or DevRel at a devtools company and want to know where you stand, [try the software](/agent-visibility) or [book a call with our team](/services). Or send us your domain and we’ll show you which tools agents pick instead of yours.

Naman