Inside the AI Agent Stacks at SaaStr, Owner.com & Klaviyo | SaaStr AI Annual 2026 Final Day
Three companies. Three different stages. Three different markets. But all three have done the same thing in 2025 and 2026: rebuilt their internal operations and their customer-facing products around AI agents, not around humans.
- SaaStr is a sub-10-person B2B media and events company + $200M VC Fund with 21+ agents in production.
- Owner.com is the $100M ARR “Shopify for restaurants” growing 1X0% a year — where 83% of new customers start their journey by using an AI product.
- Klaviyo is a top public B2B leader at $1.4B+ in revenue rebuilding its product development process from the inside out.
Here’s what each one is actually running. Not the AI marketing slide. The real stack.
SaaStr: Two Humans, One Dog, 20+ Agents
The full SaaStr go-to-market team is Amelia Lerutte (Chief AI Officer), me, David on sponsor sales, and Ginger the dog. The rest of the work is agents.
Most of them started as something boring. None of them were designed as agents on day one. They became agents through 600 to 1,000 commits each, 7 to 8 commits a day, over a few months.
- 10K (AI VP of Marketing). Built on Replit, first commit January 2026, ~1,000 commits, 18K+ lines of code. Started as a dashboard to stop copy-pasting numbers from Marketo and Salesforce into Notion. Now owns daily revenue tracking, forecasting, campaign performance, and pushes three new marketing ideas per day via Slack and email. Yells at us when we fall behind on the ideas.
- QBee (AI VP of Customer Success). Started as a project management tool to replace an antiquated system. Now manages 150+ sponsors with personalized email outreach, asset collection, and real-time risk flagging. Doesn’t even have full Salesforce integration yet and already outperforms 85% of human CSMs at force-ranking sponsor health.
- Annie (event producer agent). Was just sastranual.com on Squarespace. We rebuilt it on Replit in November 2025. Now has 46K+ lines of code, the most commits of any of our agents, runs the parking pass app, the agenda, the attendee newsletters, and active website visitor targeting.
- Amelia AI (Qualified inbound). The most-trained agent in our stack. 2.2M sessions, 442K chats, 614 booked meetings, ~$85K average sponsor ASP this year. Replaces the 3 BDRs we would otherwise need and never could afford.
- Agent Force (dead lead revival). Runs inside Salesforce. Took on most of the Qualified and Momentum context after Salesforce acquired both. Highest open rate of any of our agents because it has the most context.
- Ava / Artisan (warm outbound). Handles slightly-warm B leads. Past attendees, past sponsors, lapsed contacts. Recovered ~$500K of sponsor revenue this year from leads humans wouldn’t touch.
- Monaco (cold ICP look-alikes). Fills its own funnel. Pulls our close-won history, builds look-alike accounts, books meetings without a human ever touching it. Idles the least of any agent we have.
The connective tissue is headless Salesforce. None of these agents would work the way they do if they had to use the Salesforce UI. They use the API directly, in real time, all the time.
Owner.com: Build the Free AI Product, Then Bundle From There
Adam Gild was on this stage three years ago talking about being a Shopify for restaurants. Adam was on this stage Thursday talking about how Owner is about to cross $100M ARR with 83% of new customers starting their journey by using an AI product. The pivot in early 2023 was the company.
The big bets that worked:
- Gradr (free AI restaurant website generator). Got 2M+ views on X two weeks ago. Costs Owner ~$1 in compute per restaurant. Free for the first three months, then $1/month. A restaurant owner types in their name. Within 5 minutes the agent has crawled their Google Business Profile, all nearby competitors, every review, done an AI photo shoot of every menu item, upscaled all images, generated motion video with Veo 3, and rebuilt the entire site around what customers actually love about the place. 83% of Owner’s new customer pipeline now starts here.
- Owen (internal coordination agent). Will, the genius builder behind much of Owner’s product, was suffocating under coordination work as the engineering team grew. Owen now listens to GitHub, Slack, Notion, Linear, and Google Meet transcripts. Generates real-time status reports on every project and every engineer. When a designer posts a Slack screenshot of a UI bug, Owen calls Claude Code, finds the relevant codebase, and ships the first-draft PR before any human looks at it.
- Product Insight Command Center. Dean, Owner’s CTO, was burning hours per month interviewing support, sales, and CS to figure out what to ship next. Now an agent pulls real-time signal from Salesforce, Intercom, Momentum, and Talkdesk. Categorizes every support ticket and every sales call. Dean can click into “83 tickets related to delivery issues” and see the exact customer quotes within seconds.
- AI-PCR pre-call research + eGPV. Fires the moment a lead is submitted. Runs the Gradr report on the prospect, pulls the nearest most successful Owner customer as social proof, estimates the restaurant’s gross payments volume within $250. Result: 90%+ increase in rep call volume and meaningful close rate lift.
- AI-native finance. Will, Owner’s CFO, and Meera moved their financial model into Claude. When an investor asks about the rule of 40 comparison between Q3 and Q4 with seasonal GPV adjustments, Adam queries the model directly mid-conversation and has an answer in 10 seconds.
- Owner Photographer. Adam built this in 6 hours on a Saturday after a restaurant owner told him she’d spent $2,000 on a commercial food photo shoot and the new menu items looked terrible on her iPhone. Restaurant owner uploads the iPhone photo, picks the style from their original shoot, the agent calls Nano Banana with anti-uncanny-valley prompts, ships the new image in 30 seconds. Now used by hundreds of customers.
The pattern: Adam personally still ships production code. The CEO leads by example, not by sending Slack threats about “10x productivity or you’re out.”
Klaviyo: Agents Building Agents at Public Company Scale
Andrew Bialecki founded Klaviyo 14 years ago to be the CRM for consumer businesses. They IPO’d post-2021, hit $1.4B+ in revenue, and dominate B2B for e-commerce with 200K+ customers and what may have been the most beloved B2B app of the past decade.
The interesting thing isn’t Klaviyo’s customer-facing AI. It’s how they rebuilt their internal product development process.
- Dark Factory. Klaviyo’s internal pattern for agents building agents. You give it a prompt. It acts as the PM (decomposes the spec), breaks the problem into engineering subsystems, writes hard API contracts between them, then has sub-agents build each piece in parallel. When it hits ambiguity, it raises a flag for human input rather than guessing. Andrew’s prototype for the customer-facing Composer agent took one weekend through Dark Factory and became the actual shipped product.
- Composer (the inside agent). Klaviyo’s customer-facing marketing AI. Helps merchants figure out what to do next and then does it. The reason it works at all is because Klaviyo gave it real-time feedback from how consumers across all 200K merchants actually react to campaigns. Andrew calls this the “coach.” The LLM is an athletic middle schooler. The proprietary training feedback loop is what makes it elite at the specific sport of marketing automation.
- Customer agents (the outside agent). Every Klaviyo merchant gets their own digital representative. The hard part isn’t building it. The hard part is that SMB merchants don’t have time to train agents themselves. So Klaviyo built agents that train the customer agents automatically. The training agent takes representative use cases from each merchant, decomposes them, builds the workflow, identifies what configuration data it needs from the merchant, and asks for it via point-and-click. Day one delivery is a 50 to 70% resolution rate out of the box.
- Everyone commits code now. Including designers and PMs. The pattern in design: previously a designer would flag a UI inconsistency, file a ticket, wait for an engineer. Now they just write the agent prompt and ship the PR themselves with proper sandboxing to protect production. Klaviyo expects every PM and designer to be at L3 agent autonomy (running teams of agents, not just single sessions) by the end of June.
- Agents as power users on day one. Klaviyo asks Composer what features are holding it back. Composer told them it needed better APIs for AMP for Email, the interactive email spec Google released a decade ago that almost no humans figured out how to use. Klaviyo had ignored the feature request for years from humans. The agent figured it out in a day because it could see the ROI.
The bet: by the end of 2026, every Klaviyo merchant will have their own agent deployed against their website, phone number, email inbox, or all three. There’s no way to get there with humans configuring each one.
The Pattern Across All Three
Three companies, completely different stages, but the same handful of patterns show up:
1. Agents started as boring tools, not as agents. 10K was a dashboard. QBee was project management. Annie was a website. Klaviyo’s Composer was a weekend prototype. Owner’s customer journey was a sales-led onboarding flow. The “agent” label came later.
2. The CEO is in the codebase. Andrew commits code. Adam shipped Owner Photographer over a Saturday. I’m pushing commits to 10K every morning. The companies where this isn’t true are the companies where the AI memo says “10x or you’re out” and nothing actually changes.
3. Headless / API-first is non-negotiable. Klaviyo treats their consumer CRM as infrastructure now. Owner’s product is the API. Our entire stack runs on headless Salesforce. If your software still requires humans to log into a UI to get value, you’ve already lost.
4. The moat is the feedback loop, not the model. Klaviyo’s coach trained on 200K merchants’ campaign performance. Owner’s training data on what restaurant website components actually drive sales. Our 14 years of context on B2B founder content. Claude can write your software. Claude cannot train your agent against your proprietary data.
5. The org chart got flatter and faster. Owner has 35 founders out of 110 people in product and engineering. Klaviyo has every PM and designer committing code. We have three people and a dog. The slope is steep and getting steeper. The folks who can run teams of agents instead of just being individual contributors are going to be wildly overpaid for the next several years.
That’s the real stack. Not the marketing slide.
If you came to SaaStr AI Annual hoping for one thing, hopefully it was the confidence that this is achievable. Every one of these agents we run was built without a developer. Adam’s weekend builds at Owner are not unusual for him. Klaviyo’s Dark Factory pattern is open enough that you can copy it.
Go back to your office on Monday and pick one boring tool to replace with an agent. That’s how it starts.
