Amelia and I just released Episode #003 of The Agents. Three humans, 20+ agents, revenue went from -19% to +47% YoY, and every week we get into what’s actually working, what’s breaking, and what you should do about it if you’re running agents in production.
This week was the week the legacy stack started really showing the cracks. Our AI VP of Marketing now ships campaigns on its own. Marketo couldn’t honor its own unsubscribes for 10 days running, so we vibe-coded the fix ourselves in 20 minutes. Salesforce sent us a bill that’s 80% higher than last year even though we have fewer seats. And we realized we’d stopped using Notion months ago without telling anyone. Including Notion.
Here are the top 10 learnings from Episode #003:
#1. Our AI VP Marketing “10K” Is Now Generating, Segmenting, and Drafting Campaigns On Its Own. Seven Days a Week.
10K, our AI VP of Marketing, crossed a real line this week. It started shipping campaigns autonomously.
Here’s the actual workflow. Saturday morning, before anyone was up, 10K emailed Amelia and me with its top three campaign ideas of the day, ranked by impact, based on the latest Salesforce data, ticketing data, sponsor data, and pacing against last year. Idea #1 that morning: paid attendance is up YoY but a higher proportion of buyers are single-ticket. Run an upsell to the team-pack tier on the cohort with company size and existing seat counts that suggest they could bring more people.
Then it did all the work. Pulled the segment from our ticketing data. Cross-referenced with Salesforce. Suppressed the people the signals said wouldn’t convert. Drafted the email copy. Lined it all up with a “ready to send” status.
It stopped short of hitting send. Not because it can’t, but because Marketo’s API is so old that pushing through it autonomously is genuinely risky. (More on that below.)
This isn’t a CMO. We’re not calling it a CMO. It’s an AI VP of Marketing, and an honest one. But here’s the thing: 50% of what a great VP of Marketing does, every single day, on Saturdays and Sundays, is a lot more than 99% of companies have today.
#2. Marketo Can’t Honor Its Own Unsubscribes. We Vibe-Coded the Fix in 20 Minutes.
Day 10 and counting on this one.
Quick context: people who unsubscribe in Marketo were still receiving our newsletters. People who got fully deleted from the database were still receiving our newsletters. That is not a feature toggle issue. That is not a form sync issue. That is the platform itself failing on its single most important compliance function.
We escalated. Took until day 7 for them to even agree to investigate. Got on a call. The humans on the other end told me they hadn’t replicated the issue yet, didn’t know what was going on, and asked if I could explain it again. I asked why we got on a call.
So I went into Replit. Hooked up the Marketo API (which Replit told me hadn’t been touched since 2014). Built a custom unsubscribe page in 10 minutes that bypasses Marketo’s broken form flow entirely and just calls the API directly to delete the record. Swapped out the default unsubscribe link in our footer. Done. No more angry emails from people who got served a newsletter after asking three times to be removed.
Day 10 from Marketo. 20 minutes from us. That’s the gap.
#3. Customers Can Now Vibe Their Own Fixes to Your Bugs.
This is the meta-point, and it matters more than the Marketo story itself.
A year ago, if your vendor’s API was broken or their unsubscribe page didn’t work, you were a prisoner. Your data was hostage. You had no choice but to wait, escalate, hire a consultant, or migrate (which would take six months). Vendors knew this. It’s why support quality at most legacy SaaS platforms has been mediocre for a decade.
That trade is over.
Amelia is not an engineer. She’s a sharp operator who’s gotten technical because she had to. She vibe-coded a working fix to a CAN-SPAM compliance bug in a 20-year-old, billion-dollar marketing automation platform in less time than it took the vendor to acknowledge the bug existed.
If your customers can route around your product faster than your engineering team can fix it, you have a real problem. The vendors who survive 2026 are the ones whose pace matches the agentic pace of their customers. Everyone else is going to look slow. Then replaceable.
#4. Our Salesforce Bill Went UP 80%. With Fewer Seats. And It’s Worth It.
Two years ago we had a normal sales-team-sized Salesforce instance. Today we have two real seats and one admin seat. We’ve cut seats by 60-70%.
If pricing had stayed the same, we’d be paying less than half what we paid in 2024. Instead, our bill is up roughly 80% YoY. About $22K this year vs $12-16K last year.
Why? Because the pricing model has shifted. We’re paying for data calls and agent actions, not seats. Agentforce usage. API calls from 10K. The Slack Enterprise upgrade so we can use Slackbot agentically. We blew through every “free” Agentforce credit in days.
The “death of the seat” narrative has it half right. Yes, seats are collapsing. But they’re not collapsing into zero. They’re collapsing into consumption. And for a customer with deep agent integration like us, consumption is more expensive than seats were.
That’s not a bad deal. It’s a different deal. Net-net it’s worth it for both sides. We get vastly more out of Salesforce than we did when we had 15 seats and nobody logged in. And Salesforce gets paid for the part of their platform that’s actually getting used.
If you’re a CIO benchmarking against last year’s seat count, you’re going to misread this entirely. Watch the consumption.
#5. Headless Salesforce: I’ve Never Logged In, But I Use It Every Day.
I do not have a Salesforce login. I’m not sure I ever did. Jason at SaaStr does not exist as a Salesforce user.
But I check Salesforce every single day, because Salesforce is now headless inside 10K. Every dashboard 10K builds, every campaign idea it surfaces, every segment it pulls. All of it is Salesforce data, exposed through an API into an interface I actually use.
Salesforce announced “Headless 360” last week. We’ve been running it for six months. The official launch is the validation, not the news.
The thing most people miss about headless is the second-order effect: data that was previously locked behind a UI nobody wanted to use becomes the operating layer of the company. Our 10K is now connected to Salesforce, our newsletter platform, our ticketing data, our social analytics, and our event ops. Every metric, every attendee, every dollar, every sponsor, every speaker. All in one queryable surface that we can talk to.
I run reports against Salesforce now for real, every day. Because now I can just ask 10K (our AI VP Marketing) to do it.
#6. The Owner.com CTO Quietly Built the Same Thing. And Used It to Reverse-Engineer Every Won and Lost Deal.
Dean, the CTO at Owner (one of our portfolio companies, doing nine figures growing triple digits), built his own version of headless Salesforce on a hackathon before anyone called it headless.
What he did with it is the part worth highlighting. He took every recorded call from every prospect, won and lost, pulled the transcripts through Momentum, and used AI to analyze the actual feature gaps that lost them deals and the actual reasons they won the others. Not anecdotal. Not what the rep wrote in the CRM. The actual transcript-level signal across hundreds of calls.
His comment to me on the loom was the line worth quoting: “I didn’t used to enjoy working with Salesforce. Now we have an incredibly powerful data set that’s figuring out, in real time, why we win and why we lose.”
That is the second-order win of going headless. The data was always there. Most companies just couldn’t get to it.
#7. The N=1 App. We Built an AI Parking Pass Distributor in 90 Minutes for SaaStr Annual.
Every year, distributing parking passes for SaaStr Annual eats a full week of someone’s life. Four thousand individual PDFs in one giant file. A human had to manually split them in Acrobat, save each one, then match them to attendee requests as the requests came in by email. By the end, three of us would be working the inbox, trying not to send the same pass twice.
This year, 90 minutes total.
Step 1: I gave Claude Cowork the folder of 4,000 PDFs. Asked it to split them into individual files. Done in two minutes. Then I asked it to append the actual street address (the original passes just say “East Lot” or “South Lot,” which is useless to drivers). Done in five more minutes.
Step 2: I built a request form on Replit that takes the attendee’s info, lightly checks it against our Salesforce to make sure it’s a real ticket buyer, then routes them to the correct PDF based on attendee type and days requested. If they need two single days instead of an all-three-days pass, the agent issues two separate single-day passes and burns each one so it can never be reissued. It logs everything to a Google Sheet via a Zapier webhook as a backup.
Step 3: We launched it.
The point isn’t the parking pass app. The point is the category. These are the N=1 apps that nobody else needs but you. You can vibe-code them in an afternoon, and they replace what used to require an FTE-week of someone’s life. Every company has 20+ of these tasks sitting around. Most companies are still doing them by hand because nobody thought you could just go build the fix.
#8. Where the Hell Do You Find Another Amelia? (Aaron Levie’s New $500K-$1M Operator Role.)
Aaron Levie was on 20VC this week and called the shot: there’s going to be a whole new category of $500K-to-$1M-a-year operator jobs. Somewhat technical. Not engineers, but technical enough to deploy agents, hook up APIs, and fix what’s broken.
He’s right. He’s also understating how hard these people are to find.
We’ve been looking. Everyone we know is looking. The LinkedIn pitches you get are people who deployed one chatbot somewhere and want to call themselves an Agent Ops lead. That isn’t enough. The bar is: can you understand 20 production agents and the products they touch, well enough to spot problems before they happen, deploy new ones in days, and hold the FDE conversation with every vendor we work with.
Three personas seem to actually work:
Top FDEs at hot startups. The best Forward Deployed Engineers at Replit, Salesforce, Lovable, Vercel, etc. have already done 30+ deployments. They’ve seen what works. They’re expensive and most of them won’t leave. But they’re the closest thing to a hireable version of this role.
Ex-technical founders who are burnt out on being CEO. They’re technical, they understand the whole company surface, and they don’t need to be CEO again. The risk is burnout. They may not have another full ramp in them.
Someone homegrown internally. Probably the highest yield path for most companies. Find the smartest generalist on your team who already loves AI tools, knows what an API is, and has been quietly building things on the side. Pay them well. Give them the title.
The one criterion you cannot cut on: they have to actually love product. The mediocre version of this role looks like an Ops person managing vendors. The real version looks like a builder.
#9. The Irony of 2026: AI Companies Have Forgotten How to Do Support.
Get Recall, a tool I use every day for transcription on the SaaStr content stack, broke for about an hour last week. I went to find support. There’s no chat. No email. No help center I could find. Eventually it fixed itself.
Clerk did this to me for weeks earlier this year. Replit had no real support 12 months ago (they’ve fixed it, and the team is great now). Most of the hot AI companies I use as a customer have shipped their developer experience and skipped their support experience entirely.
The irony is loud: support is the #1 deployed agent use case across Salesforce, Replit, Lovable, and Agentforce. AI is making everyone else’s support better. AI companies have decided they don’t need to provide any.
A chatbot is not support. A “contact us” form is not support. A trained AI agent on your docs is the floor, not the ceiling. If you’re an AI vendor and your customers can’t reach a human in under an hour when something breaks, your churn risk is higher than your dashboard says. Because customers like us will just go vibe-code around you. (See learning #3.)
Best in class right now: Forward Deployed Engineers, even for SMB. Not just for enterprise. The vendors winning right now have figured out that an FDE-light model (one human, three or four customers each) is what drives adoption past the demo.
#10. Stealth Churn Is the New B2B Metric. We Quietly Stopped Using Notion Months Ago and Didn’t Notice.
The Redpoint CIO survey came out this week. 54% of CIOs are consolidating vendors. 45% of all new AI budget is being stolen from existing line items. Customer service, finance ops, project management, and sales automation are the four categories getting hit first.
Here’s the more uncomfortable version of that data, from inside our own company: I just realized we haven’t logged into Notion in months. Not Amelia. Not me. We used to run our standup there. Then 10K became the dashboard. Then 10K became the standup. We didn’t decide to leave Notion. We just stopped opening it.
We’re still paying. Notion has no idea. Their renewal will go through automatically and they’ll mark us as a happy customer in their dashboard until the day we cancel. We’re zero-DAU and zero-WAU. That’s an F on any consumer metric and we used to think DAU/WAU/MAU was a B2C concept.
It’s not anymore. In the agent era, your engagement metrics are your churn metrics, and they’re moving on a 90-day window, not an annual one. Same thing happened with Canva. Amelia hasn’t opened it in 100+ days. Same thing happened with our project management stack. We didn’t churn loudly. We stealth-churned.
If you run a B2B platform and you don’t have a dashboard tracking which paying customers haven’t logged in for 30, 60, 90 days, you do not actually know your churn rate. You know your cancellation rate. Those are different numbers now.
The vendors paying attention are watching DAU/WAU/MAU like a B2C company and reaching out the moment usage drops. The vendors still measuring success by ARR retention will be surprised in Q4 by how many of their “happy customers” already left.
That’s Episode #003.
Next on our list: building 10K all the way through to autonomous send (it’s mostly the API warmup that’s blocking us, not the agent), spinning up an AI VP of Sales / sales aura that can close $25K-$50K ACV deals end-to-end, and building the AI VP of Finance for collections that we’ve been talking about for three episodes now.
Two weeks to SaaStr AI Annual. May 12-14, SF Bay Area. Come watch us deploy agents live, run the AI VP of Marketing live on stage, and probably break a few things live too. Paid ticket attendance is well ahead of last year and sponsor revenue is up significantly.
Grab tix here.
Episode #004 drops next week.
The Agents. Every week. Three humans, 20+ agents, one real 8-figure B2B + AI company.
