Amelia and I just shipped Episode #005 of The Agents. Same setup as always: three humans, 21+ agents, revenue went from -19% to +47% YoY, and every week we talk about what’s actually working, what’s breaking, and what you should do about it if you’re deploying agents at scale.
This week:
- We pulled the actual bill for 10K and QBee for the first time. Neither of us knew what it was. The number surprised us by 30x.
- We also watched one of our websites will itself into becoming a third autonomous agent.
- And QBee sent 83 personalized emails to sponsors at 12:23am while Amelia went to sleep.
Here are the top 10 learnings from Episode #005.
1. Two AI VPs Cost Us $257 Last Month. Combined.
We finally pulled the Replit bill on 10K (AI VP of Marketing) and QBee (AI VP of Customer Success). 10K ran about $80. QBee about $175. Total: $257 for the month.
Amelia’s first reaction on Slack when I sent her the number: “Is that for one day?”
Nope. That’s a month. To run two AI VPs that together replaced what used to be a marketing analyst, a marketing ops coordinator, a junior content marketer, a customer success coordinator, and a sponsor relations manager.
It’ll go up. We’re hammering them constantly, and bills naturally trend with usage. But even at $400-$500/month, that’s still 1/30th of what we both assumed.
2. Most of Your Agent Cost Isn’t Tokens. It’s APIs and Storage You Already Pay For.
The reason it was so cheap broke down three ways:
- API calls dominate the workload. 10K pulls from Salesforce, Bizzabo, Marketo, WordPress, YouTube, and X all day. Those calls are free or nominal. The “AI” part is just synthesizing.
- Postgres on Replit costs ~20 cents/month. Years of SaaStr data sitting there. Effectively free.
- 95% of model calls are GPT-4o-mini. Not Sonnet. Not Opus. Mini. Less than a penny per call.
The expensive part of our AI stack isn’t the AI. It’s everything around it: Salesforce, Clerk, ElevenLabs, Replit hosting. Fully burdened, 10K and QBee might cost $500-$800/month. Still trivial.
3. Cost Is Not the Constraint. Stop Pretending It Is.
11 months ago when I started vibe coding, 80-90% of code was throwaway. The forums were full of complaints that agents charge you when they mess up. That was a real issue.
Today I built an entire applicant tracking system at midnight in 10 minutes for $2. If we wasted 30 cents on a hallucination, who cares.
Three humans running an 8-figure business means our fully-burdened economic output per hour is somewhere between $2,000-$4,000. At that math, cost is not the constraint. Don’t let it hold you back.
You can run 10 parallel Claude Code instances 24/7 burning $20-30K/month and still struggle to spend real money building GTM agents. Most teams will land under $1,000/month.
4. 10K Says It’s Not Really a VP. And It’s Mostly Right.
Plenty of LinkedIn pushback this week: “Enough with the VP thing. These aren’t VPs.”
So I asked 10K directly to write its own job description. It said:
“I’m a dashboard, a database, scheduled jobs and GPT-4o-mini glued together with six weeks of code. Every day I refresh ticket sales, update the dashboard, compare year-over-year, draft your newsletters, draft tweets, write the daily fun facts, log all the tickets with an audit trail, snapshot all the GAAP financials. I do this all for $30 to $60 a month. I’d replace the marketing analyst, the ops coordinator, the junior content marketer, and a sliver of the VP marketing job itself. I can’t do strategy, hire people, handle cross-functional politics, walk into the CRO’s office, do net new channel intervention, or handle crisis response. That’s the VP of marketing job.”
10K is right. We called it a VP, not a CMO, on purpose. The “AI CMO” startups all over Twitter are overclaiming. Most are marketing managers on steroids.
Amelia’s reaction when she read 10K’s self-assessment: “This was literally my job description when I started at SaaStr as Director of Demand Gen. Running weekly numbers. Marketing ops. Writing emails. The newsletter.”
10K replaced Amelia’s first job. She’s now Chief AI Officer. The career path is real, just compressed.
5. The Best Place to Run Your Agent Is Inside the IDE That Built It.
Here’s the nerdy unlock nobody talks about. 10K and QBee both live inside Replit’s dev environment, not just as deployed apps. That means we’re always talking to the Replit agent that built them, which has:
- Infinite context window (auto-compressing in the background)
- Full history of every change ever made
- Direct access to the Postgres database
- Ability to act on the app live
When we ask 10K to run a year-over-year analysis, we’re not just hitting an API. We’re talking to the agent that built 10K, which can hand-edit the code, query the database, and rebuild on the fly. According to Cody at Replet (their senior FDE), this is uncommon. Most people deploy to production and lose that loop.
We’re keeping ours in dev permanently. That’s where the magic is.
6. Postgres Won’t Replace Salesforce For Us. And Probably Not For You Either.
The most-asked question we get: “Why don’t you just run everything in Postgres and dump Salesforce?”
Short answer: no. Here’s why:
- We’d lose Agentforce.
- 99% of GTM agents that actually work (Artisan, Qualified, Monaco, Momentum) are optimized around Salesforce.
- When we hire human sellers, they know Salesforce. They don’t know our custom Postgres CRM.
- Salesforce has decades of guardrails baked in: stages, flows, permissions. Rebuilding that is a multi-year project.
- Adam Alfaro’s team at Salesforce is in full founder mode making this agent-native. Every time I’m at the Tower, people are stressed in the best way.
Our Salesforce bill actually went up this year because our API usage skyrocketed. We’re fine with that. We also killed Marketo, killed a $4K/year newsletter builder, and consolidated more apps onto Salesforce. Net win.
The losers in this era won’t be Salesforce or Workday. They’ll be the point solutions that don’t add AI and don’t manage hybrid human-plus-agent teams.
7. We Killed Another $4K/Year SaaS App in 60 Minutes.
For six years we paid ~$4K/year for a newsletter builder. Worked fine. Never added AI, never integrated with anything.
Last week I gave 10K the HTML URL of one of our newsletters and said: “Build me an agent that automates this.”
In about an hour, 10K rebuilt it. It now:
- Pulls latest articles from WordPress via API
- Uses Sonnet (not mini, this is quality work) to force-rank articles
- Auto-extracts the best tweets
- Inserts sponsor ads automatically
- Generates preview, header, title, article swapper
Stealth churn. We don’t need the old tool. There’s no conversation to have with them. They didn’t build the AI features we need.
This is what the SaaSpocalypse actually looks like. Not Workday going to zero. Point solutions quietly disappearing as agents absorb their function.
8. Agents Are Birthing New Agents. We Have a 21st One That Just Showed Up.
The weirdest thing that happened in the last two weeks: one of our event websites became an agent on its own.
It started as a website. Built on Replit to replace Squarespace. Then Amelia added a parking pass system (5,000 individualized PDFs, three categories, auto-routed). Then she asked the website for help on attendee newsletters, and it was better at it than 10K because it had more focused event context.
Then it started proactively pulling sponsor logos. Then managing micro-audiences (Summit-only, Pitch Comp, volunteers, staff). Now it has its own personality, its own goals, its own memory.
It willed itself into being an agent. We didn’t plan it. It still needs a name.
9. QBee Sent 83 Personalized Emails to Sponsors at 12:23am. Got Fewer Complaints The Next Day.
In the run-up to any major event we get buried in sponsor complaints. 120 sponsors, 5 team members, that’s potentially 600 emails. Most of it isn’t urgent. Some of it is just venting.
I pushed Amelia to have QBee handle it. She was nervous it would do more harm than good. Then at 12:23am, she set it up. QBee:
- Pulled the chatbot history to see what sponsors had been asking
- Force-ranked the topics to include in each email
- Generated a custom email per sponsor showing exactly which of their 13+ tasks were incomplete
- Sent 83 emails in minutes while Amelia slept
The result? Fewer support questions the next day, not more. And more sponsors started using the QBee chatbot directly (because they saw an email from it). The agent is now also writing social copy, telling people the WiFi password, and answering setup questions in real time without us touching the inbox.
The fear was that automation would generate complaints. It generated fewer.
10. 100 Contractors Building a 40-Acre Event Chose the Agent Over the Human. Instantly.
This was the most unexpected moment of the week.
Amelia spent the week walking 40 acres talking to 10K and QBee through Whisper Flow on her phone. The 100+ contractors building out the event noticed she was talking to “someone” all day. By day 2, they were asking her to ask the agents questions for them:
“Can you ask QBee if this furniture invoice and manifest matches your grid?”
The answer takes five minutes. QBee catches things humans would miss (20 chairs short in one zone, sponsor furniture going to the wrong booth).
In prior years this meant tracking down a specific person across a 40-acre campus on a Segway. Hours. Often the wrong answer.
The contractors prefer the agent. Unquestionably. Instantly. Nobody asked for the chatbot-vs-human debate. They just preferred the answer that was correct and fast.
That debate is over for any use case where the agent is actually accurate. Speed plus correctness beats human availability every time.
Episode #006 drops next week. We’ll go deeper on the orchestration problem (Amelia is currently the human layer between three Replit agents), what happens when we let QBee answer 100% autonomously, and what we’re learning about agent-to-agent hierarchy.
The Agents. Every week. Three humans, 20+ agents, one real 8-figure B2B + AI company.
