Hey Reader,

Last week, about halfway through my call with Joel Horwitz (founder of Synter - an agent platform for paid media), he told me about a founder who had launched one LinkedIn campaign with an agent and was happy with how it came out. The founder said that his CAC had landed around $45.

Then Joel asked him the question that I'd now ask to every founder who is running agents anywhere in their growth stack.

The question: “How much did you spend on AI to get there?”

Fact of the matter is that Joel counts three components where most startups only count one. There is your paid media cost, there is your model cost, and there is your operating cost.

Joel himself spends more on ads than he does on tokens, and he says that's deliberate.

He runs a lot of open-source models, and he told me he's fine-tuned his own model on opted-in, anonymized data. He also runs a masking layer that sits between his ad platform data and his model provider, because Google's terms wouldn't let him pass that data to Anthropic in a form that Anthropic could read.

Joel runs Synter now but before that he ran growth for Amp, the coding agent that Sourcegraph built. Since April he's written 7 million lines of code with agents, and he put roughly half of a $500,000 raise through live ad accounts to test and identify what breaks. I wanted to hear his learnings from these 5 months of testing, and he was generous with them.

READ THE FULL BREAKDOWN

What We Discussed

• Why Joel calls third-party skill files malware, and what Snyk and Vercel found when they scanned close to 4,000 agent skills across the major marketplaces
• The vocabulary that Joel decoded in about 30 seconds: tools are Python scripts, skills are markdown files, and why that reframing lets you make your own architecture decisions
• His two-agent pattern, a builder agent plus an architecture agent, and the full-funnel LinkedIn campaign that launches in under an hour because of it
• The three components of his CAC math, and why a 1 million token context window does not fix agent degradation
• The server-side conversion tracking that he built at Sourcegraph before he spent a dollar on ads, where Google Ads, Google Analytics, Google Tag Manager and PostHog are wired together and every email address is hashed
• The 2015 Apache Spark launch at IBM, why almost none of those 100 data products reached production, and the same pattern that is running again right now

Referenced

• Synter, Joel's agent platform for paid media
• Sourcegraph, where Joel ran growth for Amp before he started Synter
• skills.sh, the open agent skills directory from Vercel that Joel starts from
• Snyk and Vercel, whose joint scanning program covered close to 4,000 agent skills
• Agent Shield, the skill file scanner that Joel has open-sourced
• 1Password and Doppler, the two secrets managers that Joel named
• PostHog, one of the four tools that he wired together for server-side conversion tracking, alongside Google Ads, Google Analytics and Google Tag Manager
• Apache Spark, and IBM's 100 data products in 100 days campaign from 2015
• Lovable and Replit, the build tools behind the applications that Joel says are going to stay demos
• Model Context Protocol (MCP), and the trailing forward slash that took down half of my connections the morning of our call

Where To Find Joel Horwitz

• Joel Horwitz on LinkedIn
• Synter

I asked Joel what's actually limiting his product adoption, and he pointed straight past the technology and at the org chart instead: "I think our limitation right now is not Synter. It's change management."

Truth of matter is, I run into this limitation in nearly every engagement that I take.

Most people won't make a structural change because they could lose their job over it, so they automate the reporting and leave the operating model exactly where it was. Joel flips that fear around, and - in my opinion - his version is the correct one: "I would be scared not to make those type of changes."

If you're wiring agents into your own growth stack right now, the CAC recalculation, the secrets manager move, and the skill file scan are the three that I'd have you run this week. Each one costs you an afternoon, and together they're the difference between an agent that is measurably earning its place and one that is quietly inflating your numbers.

Joel and I are picking this up again on September 21st while he's in London, and the full founder feature will follow from that.

READ THE FULL BREAKDOWN

All the best,

Lillian Pierson

Fractional CMO & GTM Engineer