Source: https://www.anitakirkovska.com/blog/agent-led-growth/ # Building a growth engine with AI agents Nov 2, 2026 **A case study of agent-led growth: the experiments we shipped, what worked, and where we got it wrong.** For growth and marketing teams experimenting with AI agents. Growth is becoming an engineering problem. At Vellum, my job was to rebuild the playbook and give a three-person marketing team the reach of a much larger one. I think it was early February of 2026 when we got significantly stronger models and better agent harnesses that gave us what I started calling “GTM engineers in a data center.” So all of a sudden, we knew that we could throw tokens at a problem and scale our efforts. But not everything needs to scale, and not every problem needs AI. Here’s a quick look at some of our experiments, how we approached them, and where we got carried away. ## Agent-powered website This is the first step: your website, your main door to the internet, needs to be agent-powered. At Vellum, new product pages could take days or weeks, and the site’s separation from the product limited the experiments we could run. So earlier this year, around March, my CTO and I paired up and [moved from Webflow to a code-based site using coding agents](https://www.anitakirkovska.com/blog/coding-agents-made-me-churn-webflow/) in just a month. The website now lives in a GitHub repo with Sanity as the headless CMS, giving our team a way to work on pages and content with their assistants. Our website is now integrated with all of the other tools we use, so we can capture feedback and update the website from Slack: ![Slack thread inviting website feedback that Ava can turn into a pull request](https://www.anitakirkovska.com/images/agent-led-growth/slack-website-feedback.png) Website feedback in Slack becomes work the assistant can turn into a pull request. But this transition didn’t come without tradeoffs. A couple of roadblocks we hit that we knew we could handle: - Models weren’t good at frontend design. Well, it’s November 2, and now they are. Always bet that the model will get better. - Everyone can do it now. This is a blessing and a curse. We needed shared rules for where data lived, what agents could change, and how work got reviewed. Giving everyone the tools was the easy part. Helping the team use them well took a lot of education. ## The “viral articles” workflow We used a reusable `SKILL.md` file to produce comparison articles and benchmark explainers around new model releases. I call these our “viral articles.” Some reached tens of thousands of views in a single day. The GEO writing skill defined the research and writing process: gather live sources, save research notes, back up comparisons with citations, and run a quality check before review. We used it across head-to-head comparisons, listicles, and other formats. We also expanded beyond our core topics into adjacent categories, including voice agents, search tools, and web scrapers. Those posts created another growth opportunity: companies wanted links from our articles and offered backlinks in return. [See the GEO writing skill](https://github.com/vellum-ai/vellum-assistant/blob/main/skills/geo-writing/SKILL.md) ## Organic (AI) UGC We chased scale with both human creators and AI influencers before figuring out what worked. We wanted to move fast, and our creators didn’t get properly onboarded and didn’t fall in love with the product. That is a recipe for disaster. And our biggest mistake. If I’m starting a new UGC program from scratch, I’d start with a few handpicked creators and invest in their product knowledge, voice, and relationship with the brand. I will keep everything in-house. Obsession is key here. [Hector Serrano’s Wispr Flow playbook](https://www.lennysnewsletter.com/p/how-a-19-year-old-intern-built-the) is the example I’d study: start with a small group, invest in coaching and relationships, then scale what works. If you’re selling a technical product, you should probably go with real creators, and you can automate their scripts and post-production with agents. They can also clone themselves with tools like VEED. Higgsfield is also great for creating variations with their AI influencers on videos you shoot yourself. Here’s me talking about a new product, AI glasses, but with a totally different AI persona that kinda looks real: [Video](https://www.anitakirkovska.com/images/agent-led-growth/higgsfield-ai-influencer.mp4) An original video I shot, reworked with a different AI persona in Higgsfield. ## Content at scale Now, these are the things you can do at scale. Claude is getting so good at creating explainer videos, you just need to have a pretty good skill on the format. You can also clone your voice in ElevenLabs by passing a 10-minute video of you talking, and that’s it. You can start to create a lot of these assets. Creation will be easy, building distribution could be challenging. I’d start with shorts on TikTok and YouTube; the algorithms are generous with new accounts and you can get more engagement right away. My explainer workflow starts with a written idea and ends with a video and an article. I use ElevenLabs to clone my voice, then Claude to generate the videos. From there, the workflow automatically publishes to AIExplained and YouTube and generates a repurposed article. Three examples: - [Turning a post into a 2-minute explainer.](https://x.com/i/status/2109866179619618936) I started with: “hey Claude, turn this post into a 2-min explainer video.” - [A filing explainer.](https://x.com/i/status/2114761109064310962) I asked Claude to read a filing and turn it into a video. - [A 40-second product video for Ambient.](https://x.com/i/status/2103263314453123097) An experiment in generating a product video from a single prompt. [AI explainer video](https://www.youtube-nocookie.com/embed/77aTqr6T5gk?start=2) An example from my AI explainer workflow. [Watch on YouTube.](https://www.youtube.com/watch?v=77aTqr6T5gk&t=2s) ## Performance marketing that adapts to the user [![Five Meta ad themes lead to matching Codex-built website pages, personalized onboarding, the app, and a retargeting loop](https://www.anitakirkovska.com/images/agent-led-growth/personalized-marketing-loop.png)](https://www.anitakirkovska.com/images/agent-led-growth/personalized-marketing-loop.png) Ad intent carries through the landing page and onboarding into the product. Open the diagram to view it at full size. Above is a very complicated-looking workflow, but at its core, it’s just a personalized loop that turns a cold lead into a user as your agent learns what they need. It’s all semantics now! Your users don’t adapt to your product; instead, your agent adapts to each and every one of them. We tested a lot of AI-generated creatives on Meta, then created custom landing pages with Codex for each of the major themes, and made sure all of those data points were captured in the onboarding so that the agent knows how to activate each user. We ran ad creatives that I call “ugly ads.” Those are the ads that don’t look branded and look very native to the platform you’re running them on, like these for Instagram Stories: ![Three Vellum ad creatives about personal finances, workouts, and family logistics](https://www.anitakirkovska.com/images/agent-led-growth/meta-ad-examples.png) Different creative angles make the assistant’s value concrete for different audiences. ## Agent-led activation I wrote a [full case study](https://www.anitakirkovska.com/blog/the-product-onboards-to-you-now/) on how we redesigned our onboarding in Vellum. This is exactly where the agent now onboards to your user, and not the other way around. This experiment improved onboarding completion by 10.7 percentage points, and our day-one activation increased from 27.5% to 37.4%. ![Onboarding combines assistant setup and background research before confirming findings and suggesting tasks](https://www.anitakirkovska.com/images/agent-led-growth/personalized-onboarding.png) The assistant learns about the user while they set up its name, personality, and first integration. We used Claude Code for much of the build. The flow combined four pieces: - You pick your assistant’s face, name, and personality, while a deep research agent fills in the rest about you in the background. - You confirm what the research agent found, and the assistant suggests tasks tailored to you. - A day-two check-in. After connecting Google Calendar, users could confirm a meeting with their assistant for the next day. - A personalized email sequence. I built the onboarding drip with [my assistant, Resend, and Claude Code](https://x.com/anitakirkovska/status/2063773079369642026). I was really excited that almost 6% of users who connected Google Calendar came back through the invite because we made the assistant book a 1:1 with them: ![Calendar invitation to a day-two check-in with Anita’s AI assistant](https://www.anitakirkovska.com/images/agent-led-growth/calendar-check-in.png) A calendar check-in gives the user a reason and a direct link to return to the app. ## Building resources agents can cite The [Vellum LLM Leaderboard](https://www.vellum.ai/llm-leaderboard?utm_source=blog&utm_medium=article&utm_campaign=agent-led-growth) helped people compare models through public benchmarks, speed, latency, and price. It was an interactive resource with an ongoing maintenance job behind it. This resource was important for two reasons: - It helped us build topical authority in a field we were actively selling into: evaluations. - It helped us build relevance for agent browser searches, so we got a lot of authority and recommendations. Benchmarks and other types of evidence data are what agents crave right now. ![Vellum LLM Leaderboard comparing models through benchmark charts and tables](https://www.anitakirkovska.com/images/agent-led-growth/llm-leaderboard.jpg) The leaderboard combines public benchmark data, model comparisons, and a copyable page. We fully automated this resource: background workers checked for new models and benchmarks, then updated the page and its “last updated” date almost as soon as they launched. This page alone brought in half of our traffic per year!