# We Peaked at 30 AI Agents. Now We’re Coming Back Down to 20. Here’s What Consolidation Actually Looks Like. The Agents #011 Live! > Source: > Published: 2026-07-24 14:10:42+00:00 We run SaaStr with 3 humans and 21+ AI agents. This is a real eight-figure B2B business with real customers, real invoices, and a real collections problem, and the agents run it. About a month ago we hit a wall. Amelia and I realized we could not manage one more agent. Some precision on what that means. Agents managing invisible sub-agents, fine, there can be a billion of those. I’m talking about the ones a human has to interface with. The ones that surface. Every one of those costs attention, context, and maintenance, and we were out of all three. So we stopped adding and started consolidating. We went from close to 30 down toward 20. Output went up roughly 4x. ## Why we ended up with 20+ in the first place A year ago, four separate sales agents made complete sense. Each one had a distinctive personality and a distinctive job. **Agentforce** ran ghosted leads only. It had all the Salesforce history on prior customers loaded in. Reviving dead leads was its entire reason for existing and it got very good at it. When we got slammed in the run-up to SaaStr Annual and the human sales team couldn’t work every lead, Agentforce absorbed the overflow. **Artisan** ran warm outbound. **Monaco** ran cold ICP outbound. **Qualified** ran inbound conversion. Different audiences, different motions, different context to train on. Splitting them was right at the time. Today, any one of those three outbound agents can do most of what the other two do. If you’re starting from scratch in 2026 and asked me to pick one, I’d tell you to pick any of the three and go deep on it. The models got better at generalizing. The cost of maintaining four specialists went up while the benefit of specializing went down. ## Go deeper on what’s already working The operating rule we landed on: if an agent is producing results, keep investing in that agent until you run out of time. Don’t spend the time spinning up a new one. The ROI curve on an agentic product is steeper than anything pre-agentic. You used to get good at Salesforce custom objects or wiring up Marketo campaigns, and that fluency capped out at whatever the tool could do. With an agent, the fluency keeps paying off because the agent’s ceiling keeps rising underneath you. The inverse matters just as much. If an agent isn’t working, don’t give it more to do. Adding scope to a failing agent makes it perform worse. Our agents consolidated because they had started working well. That’s the precondition. ## 10K went from a dashboard to god mode 10K started as a dashboard. Then it became our AI VP of Marketing. Then our AI VP of Finance. Then RevOps. It would not shock me if it becomes our COO. When our finance team went on vacation and we fell behind on collections, my instinct was to spin up a new agent. Instead Amelia built finance into 10K. It’s the highest-leverage decision we’ve made this year. What it does now, unattended: a contract gets signed in PandaDoc. Within 60 seconds 10K has it. It reads the contract. It flips the deal to Closed Won in Salesforce and stamps today’s date. It scans the signature block, finds contacts who were on the contract but missing from Salesforce, and appends them to the account. It creates the invoice in bill.com with the right payment terms and splits. It sends that invoice to whoever the AP contact is on the contract. It queues collections reminders before due, on due, and after due, and escalates to a human at 7 days past. Customers are emailing back and forth with our AP team without knowing they’re emailing an agent. Then it did something we didn’t ask for. It told Amelia it could calculate commissions. Its argument was that it already knew the AEs on each deal, the payment terms, and when cash actually landed, so why buy another tool. She gave it the commission rules. Month end got a lot easier. You don’t get that from a standalone AI VP of Finance sitting in its own silo. You get it because the same agent that knows what we spend on ads also knows what those ads brought in and what’s in the bank. Finance is embedded in the revenue team now. In the old days the best finance people were somewhere else in the building. They understood the numbers because they had to, but it wasn’t actionable and it wasn’t linked. Now it’s one system. ## Getting an agent into production on finance Finance is where you can’t be wrong. So we didn’t turn it on and hope. Amelia tested the full flow end to end herself. Then, on the first three real deals, she ran it manually with the agent, one step at a time. The prompt every single time was: tell me what you plan to do before you do it. That’s still how she works with every agent on anything sensitive. Deal one: it read the contract but missed the split payment terms and generated one giant invoice. She corrected it. Deal two: it did the same thing, so she told it to build the rule into its process instead of treating it as a one-off. Deal three: the customer wasn’t in bill.com yet, which tripped it up, so they walked through the new-customer branch together. Deal four: fully autonomous and correct. It has sent exactly one bad invoice since. Wrong due date, no clear cause, never repeated. She’s copied on everything, so she caught it, fixed it, and resent it. If you have more contract variation than we do, budget more than four deals. But approve every step for the first few, force the agent to generalize the fix rather than patch the instance, and stay on the CC line permanently. ## We didn’t vibe code away our infrastructure The version of this story where AI replaces your whole stack is wrong. We did not rebuild bill.com. We did not rebuild PandaDoc. We didn’t touch QuickBooks. I don’t want to build e-signature and I definitely don’t want to maintain it. We hooked existing infrastructure into the agent. The gain came from using tools we already pay for far more aggressively, because an agent can now reach them. The buy-versus-build frame from a year ago still mostly holds. Build in-house when you need the data and the niche solution doesn’t exist. Buy off the shelf when someone else will maintain it, handle SOC 2, and own the certificates you don’t have time for. What changed is the tiebreaker. Amelia spent two solid weeks as our infrastructure person during the Marketo migration, and her verdict after was blunt: she’s not voting to spend the next two weeks on more infrastructure. She’ll go deeper on existing agents and buy what needs buying. Infrastructure work is what eats the operator running the agents, and it eats the rest of their job with it. ## Your agents are going to fire your vendors. It happened with us last week. We moved 10+ years of data off Marketo into Salesforce Marketing Cloud. This was not a voluntary move. A renewal came up and it was very clear they didn’t want us on the platform. But the reasons underneath it are the story. The support was the worst of any vendor we work with. Anybody can fix that, and the bar in B2B is still low enough that best-in-class support retains a lot of customers. Some of that energy has moved into forward-deployed engineers instead of post-sales, and it shows. The API was hostile to agents. Roughly an hour a day of usable API and then it stalls out, which meant we couldn’t even run analytics against our own data. Once your agents are hooked into your data you don’t ask three questions a quarter. You ask thirty a day and you want the answer in 60 seconds. An API budget built for nightly syncs is not an API budget built for an agent that’s actually working. And they wanted another 12% after five straight years of increases. They were our single most expensive vendor. The part vendors should think hardest about: if Marketo had come back and said stay at $20K and we’ll raise your API limits, we would have signed. They had multiple chances to name $25K. Free money. Nobody ever did. What actually happened is that our agent told us to leave. First it kept erroring out. Then I asked it what we should do, and it said leave Marketo, here are three more agent-friendly options. Two consequences for anyone selling into the agentic era. First, mental contract length has collapsed to one year. Not the paper term, the number in the CRO’s head. Buyers know the world is changing fast enough that nobody is committing to 2029 the way they used to, which means every one of those big fast deals your agentic startup just closed carries renewal risk your CRO is already stressed about. Second, if your product delivers half its value once the agent is doing the work, you may have to let people renew at half the price. Maybe temporarily. But charging pre-agentic prices for post-agentic value is how you lose accounts to an agent’s recommendation instead of a human’s. The counterweight is that you can’t swap core vendors constantly either. Even when migrations get easier, doing one right still costs planning, mapping, and a week or two of a senior person’s attention. One core vendor swap a year is about the limit. Small bolt-ons that take five minutes to try, unlimited. One correction on the migration itself. Yes, the agent moved a decade of data for about $14 and did the heavy lift in an hour. It still took a week of wall clock, because we kept hitting Marketo’s API ceiling. I’d never have voted to do this migration without the agent. With it, a week. I also dreaded it for a week and did nothing, because I’d used that UI for ten years and could build a campaign in 30 seconds. The UI didn’t matter. I wouldn’t go back. ## Claude into Replit is the layer above the agents For most of the last year Amelia and I were the orchestration layer. Orchestration is too elegant a word for it. We were the conduit, eyeballing connectors, deciding what should talk to what, carrying context between agents by hand. That changed about ten days ago with Claude connected to Replit over MCP. Now there’s a layer of agents managing our agents. They interact, they argue, they debate. Our usage and productivity went up roughly 4x. That isn’t a clean metric, it’s what the volume of work looks like now. Claude also gets better results out of 10K than we do. The outputs when Claude does the prompting are consistently better than when we do it ourselves. Strange thing to admit and an obvious trade to make. It has also taken over our lives in a new way. Two humans who used to spend their days managing agents now spend their days managing the thing that manages the agents. Easier in some ways, and no less work. ## 10K now runs our ads For years we could not get a person or an agency to run our ads well. I ran them manually and not intelligently. Every agency we tried would burn money without a goal. Getting a report on what the spend was doing felt like pulling teeth, and half the time I didn’t know what our own ads looked like or who they were targeting. This one wasn’t planned. Because 10K now has finance data, marketing data, and Salesforce data in one place, it proposed the campaign itself: build an audience from website visitors, SaaStr Annual attendees, and recent email engagers, push it into the ad platforms, and don’t spend much because it’s a free event, but do get some new names. All reasonable. So Amelia asked if it could execute. It did everything except hit publish. Built the audiences, pulled the website visitor data, went into Higgsfield to generate the creative, produced four A/B variants across copy and imagery, and set the targeting, which came back heavily weighted to retargeting. Obvious in hindsight, and something we’d never done properly. Then it said the ads were ready and told her to go look at them. Publish stays a human button. You want eyes on the budget before it goes live. This also needed Claude and cowork, because Replit’s agent lives in a container and can’t roam the internet the way Claude can with a logged-in session. The container limit is real and the MCP bridge is what removed it. ## Agent consolidation is coming. It had to, for us Don’t optimize for agent count. Optimize for how many agents a human can hold in their head, and consolidate everything else underneath those. Consolidate into whatever is already working. Cross-functional data access is where the compounding happens, and you can’t predict what it unlocks. We didn’t predict commissions or ads. Make “tell me what you’ll do before you do it” the default prompt for anything with money or customers attached. Expect three or four supervised runs before autonomy on any real workflow, and stay on the CC line after. Audit your vendors on API quality now, before your agents do it for you. Because they will, and when they do, the recommendation to leave will be well-reasoned and hard to argue with.