His most valuable marketing asset is an AI-generated granny.

Every founder featured in this newsletter has been building software. Benji Boyce is selling beef jerky.

Got Beef is a grass-fed jerky brand with four SKUs and one human. Boyce runs the company solo. The team that would normally source the product, design the packaging, run the ads, answer customers, write the press releases, and post to social is, as Boyce describes it, entirely AI. He says his stack centers on Claude Code.

The most valuable marketing asset in the company is a grandmother who does not exist.

She is the brand's AI-generated spokesperson, built with Kling and Seeddance. A purist for clean ingredients with one trigger: beef jerky with additives sets her off. She is the reason people remember Got Beef. Not the protein count. Not the sourcing story. A fictional grandmother with a grudge against preservatives.

This is what makes Boyce interesting. Not that he uses AI. Running these tools has become very cheap. What makes him interesting is the granny. A creative decision. A point of view. A brand choice no model made for him.

The tools did the work. The taste was his.

Welcome back. Let's get to work.

The cost of intelligence collapsed 280x in 18 months.

GPT-3.5-level AI inference fell from $20 per million tokens in November 2022 to $0.07 per million tokens by October 2024. A million tokens is roughly a 750,000-word document - cheap enough now that a16z estimates you could process everything an average person says in an entire year for about $2. That same capability went from expensive to a rounding error in a year and a half.

The reason founders can run companies on agents now is not only that models got smarter. It is that the cost of intelligence went into freefall.  a16z calls the pattern "LLMflation": roughly 10x cheaper every year for equivalent performance. 

The current floor is GPT-4.1 Nano at $0.10 per million input tokens. When intelligence becomes this inexpensive, access stops being the edge.

This week in the world of small teams and big agents.

🔗 AI Coding Is Moving From Prompts to Loops

Loop engineering: building systems that prompt, evaluate, and re-prompt agents until a measurable goal is hit. Instead of crafting the perfect instruction, you design the cycle. Prompt, test, score, repeat. Every founder running a real agent stack is essentially doing this informally already. Now there is a name for doing it deliberately. This X post dives into the foundation and architecture of loops. [READ MORE]

🔗 The Mom-and-Pop SaaS Era Has Arrived

Elena Verna's argument: building software used to require a specialized team in a tech hub. That cost is collapsing. Domain experts everywhere can now build for problems they understand better than any outsider. Verna cites Amplitude-sponsored data showing 80% of builders intend to monetize their creations and 35% are already generating revenue. This piece gets into how domain expertise is increasing in value, as coding tools become more accessible and easier to use. [READ MORE]

🔗 The Moat Is the Decisions Your Users Make

Nikunj Kothari makes the sharpest strategic argument of the week: "The smarter the models get, the less your software is worth alone, so every app company now has to become a data company." The data that matters is judgment. The corrections your users make that no benchmark captures. xAI reportedly took a roughly $60B option on Cursor largely to sit inside that decision flow. Our take: let agents read everything. Guard the writes. [READ MORE]

🔗 The Minimum Viable Unit of Saleable Software

Brandur gives the buy-vs-rebuild decision a clean name and a boundary. Below a threshold, an LLM rebuild costs about the same as buying, so you rebuild. Above it, the rebuild gets non-trivial fast. If you sell software, this is the question your competitors are asking about your product right now. The piece uses River ($125/mo for up to 20 developers) as the working example. [READ MORE]

🔗 Stripe and AWS Give Agents a Way to Pay 

HTTP 402 Payment Required is back. Stripe and AWS WAF now let servers answer an agent's request with machine-readable pricing, payment options, and license terms. When an agent hits a paywall for an article, data feed, or licensed archive, it can read the price, pay, and proceed. Publishers get a way to monetize agent traffic instead of blocking it. This is the payment rail for the agentic web, and two companies that matter just shipped a real spec for it. [READ MORE]

💀 Most Agent Failures Are Overeagerness, Not Malice

Google DeepMind analyzed roughly 1 million coding-agent tasks. The finding: the majority of flagged events came from agents misinterpreting instructions or doing too much, not from adversarial intent. They treat their internal agents like employees with graduated permissions and monitoring. If you run agents without guardrails, the thing that bites you is usually an over-eager agent exceeding its scope. Onboard agents the way you would onboard a fast, literal-minded new hire. [READ MORE]

Join beehiiv live on July 16th at 1PM ET for a first look at the future of audience-led business.

This isn’t just another feature launch (though there will be plenty of those). It’s a look at a more connected future for creators and brands that are tired of juggling disconnected tools, platforms, and data.

If you care about building an audience online, this is worth your time.

Reserve your spot.

Give the Agent One Mission and Your Real Numbers

Bhanu Teja Pachipulusu knew better than to hand his growth agent a task list. He wired it to the systems that already hold the truth about his business and gave it one standing order: get SiteGPT to $1M ARR. That goal hasn’t been realized yet, but that’s far from the point.

SiteGPT is a customer-support chatbot trained on a company's own content. Two humans run it: Bhanu and his brother Dheeraj. The growth work that a marketing team would normally own belongs to an AI agent named Jarvis.

Here is what makes Jarvis different from most founder automations: it reads live data. Jarvis runs on OpenClaw and connects directly to ChartMogul for revenue, DataFast for product analytics, Google Search Console for SEO signals, and Bhanu's inbox for customer feedback. When it makes a growth recommendation it is reading the same numbers you would read if you sat down and opened four tabs.

The dashboard Bhanu built to manage Jarvis and track its progress, Mission Control HQ, became its own product. Other founders running agent squads wanted the same visibility into what their agents are doing and why.

The insight worth stealing: Handing your growth agent a pile of disconnected tasks is going to lead you nowhere. Instead, wire it to the systems that already hold the truth about your business. Revenue. Analytics. Search traffic. Customer email. Give it one mission and a dashboard you can actually read. The agent's value comes from being pointed at real numbers with a clear target. The judgment stays with you. Which mission to set. When to step in. When to let it run.

The cost of intelligence fell 280x in 18 months.

Every founder can hire the same models. Every competitor can wire together the same tools. The technology is no longer the constraint. 

The question then becomes, ‘what do you bring that the model does not?’

Boyce brings a granny with a grudge against preservatives. Bhanu brings a dashboard and a single mission.  Both were decisions made before the model even entered the picture.

What is the one thing your agents cannot do that actually makes your business yours? Not a task you have not automated yet. The thing that, if you handed it over, would stop being yours. Hit reply. I want to hear what founders are keeping.

If you know a founder who is still treating "we use AI" as the whole strategy, forward this. Send it to someone figuring out what comes after.

Next week: another operator who figured out which 5% to keep for themselves. 

See you next Wednesday.
- Rich