Amelia and I just published Episode #016 of The Agents. Same setup as always: three humans, 21+ agents in production, and a B2B + AI business on track for about $20M this year.
This week we went deep on truly going “headless” and agent-first with your CRM. Amelia connected our actual Salesforce org to three AI tools: Claudeforce (just out in beta), Replit, and Lovable. Same data, same prompts. On factual questions, all three gave us essentially the same answers. The bills ranged from about $2 to $15 to set up. But when we asked each one what we should do next, we got three different lists.
The second half of the episode was about agents choosing our vendors. Our Connect agent picked a hosting company I had never used. Our SaaStr.ai agent told us which CRM it would pick if we were starting over, and told us to stay where we are. 10K told us to skip buying an AEO tool and let it build one. And two of our agents set up their own channel to talk to each other.
Here are the top 10 learnings from Episode #016.
#1. Three “Headless” Tools on the Same Salesforce Org. The Same Answers, But Very Different Value Propositions #
We’ve run Salesforce headless through 10K for most of a year, with hundreds of hours of iteration and 20+ connected apps, so putting 10K in this comparison wouldn’t be fair. Amelia left it out and started fresh on all three platforms to see what a founder gets today, out of the box.
Claudeforce is Claude’s new Salesforce connector. It’s in beta, and you request access from inside Claude. It connects over MCP and ships with roughly 35 pre-built skills, which give a sales team a starting list of things to ask.
Replit and Lovable both have built-in Salesforce connectors today. You type “connect to Salesforce” as your first prompt and it’s done. Both use the standard Salesforce API, the same way 10K does.
Amelia gave all three the same prompts: build me a dashboard and an account map, then give me the top five things we should do. Then she asked each about 10 identical questions. Who owns the Finn account? All three said Amelia, which is correct. How many deals does Amelia have versus David? All three got that right too.
All three also have read and write access. Claudeforce is restricting itself to read-only during the beta, even though it has a write tool. Replit and Lovable default to read-only and confirm with you in writing before they write anything back to Salesforce.
-> Claudeforce, for real. It’s kinda awesome, better than the ‘native’ Salesforce out of the box, but for now appears to mainly limited to Salesforce data:
#2. The Same Build Cost $2.33 on Lovable, $11.81 on Replit, and About $15 on Claudeforce #
- Lovable: $2.33. You can’t pick models in Lovable the way you can in Replit, and when Amelia asked which model it used, it couldn’t give her a straight answer. It’s clearly choosing cheaper ones.
- Replit: $11.81. Amelia started on the free tier, then the agent recommended finishing in Max mode on Opus, so she switched. Amelia and I pick frontier models in Replit on purpose, which is a big part of why 10K lives there and why our Replit bills run higher.
- Claudeforce: about $15. This is an estimate, since it’s covered by Amelia’s $200/month Max plan. She asked the agent what the session cost and it said $15 to $20 of her plan.
None of this includes Salesforce itself. Nobody knows yet what Salesforce will charge for API calls versus MCP calls at GA, including Salesforce. If customers have to pay for Claudeforce on top of a Salesforce contract to see their own data in a different view, a lot of them will be frustrated.
All were disruptively cheap at least to get going and start. But it was interesting Lovable had optimized its Salesforce connector to do it for a fraction of the cost. If that delta holds at scale, it would be material.
#3. When We Asked What to Do Next, Each Tool Gave Us a Different Top Five #
Same pipeline, three different plans for the day.
Claudeforce’s top five:
- Close the deals sitting in stage 4
- Engage on our largest diamond renewal (Google Cloud, sitting in a Q4 close date)
- Run a “blitz” on the $1M+ in stage 2 to get those deals to stage 3 or closed
- Escalate the bigger deals that have been dragging
- Push Salesforce for a firmer timeline and proposal
Replit’s top five surfaced many of the same deals in a different order. It flagged stage 3 deals with no next step recorded in Salesforce, and it was the only one to bring up negotiation terms: old blockers on decision dates, payment terms, anything in the renewal language that could stall a close. It put closing the stage 4 deals at #5, where Claudeforce had it at #1.
-> Replit’s Headless Salesforce: powerful, best at managing deals, but most work to configure:
Lovable’s #1 didn’t appear on the other two lists: rescue the overdue deals, meaning anything past its forecasted close date. That keeps the pipeline honest, and Amelia said it’s the first thing she’d do walking into Salesforce in the morning. After that: book meetings on five overdue deals in the working stage, unpark seven renewals, put dollar amounts on deals that have none, and get Google Cloud over the line.
Amelia would act on any of the three. The model and vendor behind the agent decide which list your reps get each morning, so we’d run the same prompt across two or three tools before standardizing a sales team on one.
#4. Lovable Built a Login Without Being Asked, and It’s the Only Dashboard Of the Three I’d Use As “CEO” #
Lovable was the only one to build a login unprompted. It named the app Ledger and laid it out like a newspaper. Open pipeline, about $5M, was the headline, with win rate and average deal size next to it and the top five actions above the fold. It was also the only one to show our top accounts.
-> Lovable’s Headless Salesforce: clean, built in SSO (built as an app), and the simplest to process for a “CEO”:
Amelia ranked the design Lovable first, Claudeforce second, Replit third. Replit’s first version was barely legible, and it took more prompting to get to a usable dark blue and green UI. Replit also put the actions at the top and the dashboard at the bottom.
Amelia, David, and our sales team work in the pipeline every day. I don’t. I don’t log into Salesforce, and I only go into 10K when I need an answer, like the year-over-year metrics I pulled this morning. 10K is excellent and also dense. Lovable’s version showed me in one screen what’s on the table, our win rate, our deal size, and the top things to do. It’s the only one of the three I’d open on my own.
My guess is Lovable assumed we were building an app other people would use. That also puts it furthest along on multiplayer. Most of our Replit agents are internal tools for one team. As companies roll agents out across bigger teams, every stakeholder needs a version they can log into. Claudeforce, at least in beta, feels built for one rep looking at their own data.
#5. The Benchmark Is Classic Salesforce, and All Three Beat It #
Amelia started out comparing the three tools against each other. Halfway through, she pulled up Salesforce as most companies still use it: the default seller home and a dashboard nobody configured. She hadn’t looked at that screen in about a year, and she added it to the comparison.
Each of the three was “Headless Salesforce” agents built in 20 minutes to an hour, and each is an order of magnitude more useful than logging into classic, native Salesforce.
- You can talk to them
- They analyze your pipeline, and
- They tell you what to actually work on.
- And Replit and Lovable can easily integrate dozens of your other third party apps. ClaudeForce can too indirectly, leveraging whatever connectors you have in Claude, but TBD if that’s as useful.
I saw the Claudeforce demos at Dreamforce and at the time … wasn’t that impressed from the mere demos. Getting text back about my pipeline inside Claude didn’t seem like much. But I was blown away by using Claudeforce for real inside our real Salesforce org. I don’t know anyone … wouldn’t. It looks like what people pictured the first time they bought a CRM it would actually be: one great screen, the hottest deals, and an AI telling you where to spend the day.
I give Claudeforce a lot more credit now. It will need iterations to do everything everyone wants it to do, and I’m not sure all of the pre-built skills will be intuitive for most sales teams. Once it’s GA, I don’t see why a Salesforce rep would log into native Salesforce.
#6. Claudeforce Sees One System. Our Stack Is 20+ Apps. #
For one person working only on Salesforce data, the three are roughly equivalent. Claudeforce should also get easier than the other two over time, since it’s native. When we first hooked Salesforce up to Replit, we fought expiring tokens for weeks, and a native connector shouldn’t have that problem. The key difference showed up when we went headless with all 3. Within weeks we’d connected QuickBooks, Bill.com, Agentforce, and more to 10K. 10K touched more than 30 apps at one point and touches 20+ today. Claudeforce can’t bring those in, at least not directly, so our first integration outside Salesforce would send us to Replit or Lovable, at least for now.
Small teams wanting to build an AI VP Revenue like 10K may outgrow Claudeforce quickly, because most of 10K’s value comes from combining systems. But large enterprises may pick Claudeforce because it can’t connect to everything. With a thousand sellers, you don’t want all of them writing back to Salesforce through tools you can’t lock down, and Replit and Lovable will be harder to lock down at that scale. And be more work to maintain. But all three are again … the CRM we always really wanted.
These products also still look like the people who built them:
- Replit’s output was the most developer-minded: all black, the right ideas at the top, less attention to the dashboard. That’s Amjad and a company with developer roots.
- Claudeforce looks like it was built by Salesforce people who know the app cold and rethought what it should do.
- Lovable’s layout reflects Anton’s design background: the right three metrics, sized and weighted correctly, in the right order.
We see the same thing in AI SDRs. Monaco looks like the outbound teams Sam Blond built at Brex. Artisan started from Jasper’s view that you don’t need humans, and has since backed off that a bit. These roadmaps will keep following their founders’ instincts for the next 12 to 24 months, so we spend time understanding the founders behind a tool before we commit to it.
#7. Our Connect Agent Picked Render, a Host I Had Never Used #
We’re building SaaStr AIConnect, an agentic recruiter that helps founders find strong passive candidates who aren’t in market. It’s built on Replit. I wanted to know what we’d use if we added a second platform at scale, so I asked the Connect agent live during a 20VC recording.
It listed what Connect needs: an always-on web server for the site and API, plus an always-on background worker. The Connect agents run 24 hours a day, and some jobs take 10 minutes, like analyzing 25,000 executives against one new role. That means long-running jobs, scheduled jobs, health checks, zero-downtime deploys, and longer log retention.
It picked Render. Railway was #2, very easy to use but a bit less mature for heavy production. It called AWS and Google Cloud overkill, and ruled out Vercel for minutes-long jobs and background workers. I’m not sure that’s right. We’ve used Vercel on other apps, and it’s powerful and cheap.
I follow Render on X, but I’d never used it and wouldn’t have thought of it for Connect. The agent didn’t suggest testing two or three. It named one as the clear winner and its pick.
I asked Amelia which she’d use. Render, because the agent said so. She’s not an engineer and wouldn’t test the runner-up, and most buyers in her position wouldn’t either.
Next comes the agent asking “Would you like me to move you to Render?” and you saying yes. By 2027 it will say “I’ve already migrated you, want to see it?”
I ran the same exercise recently with the founder of a Fund portfolio company. I asked 10 different agents which vendor to choose in his category, and all 10 picked his company. I told him to have someone on the marketing team check those answers every week, because the agents keep changing.
#8. Our SaaStr.ai Agent Told Us to Stay on Salesforce, and Not to Switch Without a Clear Reason #
We asked the agent running SaaStr.ai which CRM we should use, just to see what it would say. It knows all our agents run on Salesforce, and it answered with that in mind.
It said stick with Salesforce. Keep it for revenue operations, don’t migrate without a clear reason, and run a small Attio pilot for the SaaStr Fund against Affinity. Affinity was the only fund-specific CRM it named, which is a win for them.
Then it added that if we were starting from scratch today, it would look at HubSpot for the operating business and Attio for the fund. Attio honestly would never have occurred to us. It was built pre-AI, and while it has a lot of strong customers, it just wasn’t on our radar. And we’re not sure HubSpot would be the better choice for us at all. It’s no cheaper anymore (it used to be), and most of our third-party agents either don’t run on HubSpot or don’t run on it as well. The agent’s reasoning was mostly that it rated HubSpot’s marketing platform highly, and that SaaStr AI does a ton of marketing (which is true), more marketing than “sales” activities.
Whether the agent is right matters less than the fact that it had a view, and a new founder asking the same question would get that view.
The agent also made a call a human SDR would never pitch: don’t switch without a reason. We get endless outbound emails trying to move us from one vendor to another. More of those decisions will run through agents that can see your data, and some of them will say stay put.
AEO and GEO are about ranking when a human asks ChatGPT or Claude what to buy, with one chat and a small slice of memory as context. An agent working inside your company has the whole application as context, and it has opinions. I’m calling this Agent-Led Growth, or ALG. Developer tools already work this way. Clerk and Resend took off when the coding agents started recommending them. In the early 10K days I had to fight the agent constantly to use Resend instead of SendGrid. Today the agents recommend Resend on their own. We expect the same shift in every B2B category.
#9. 10K Told Us to Skip the AEO Tools and Built Its Own #
Amelia asked 10K what we should use for AEO. We want agents to recommend SaaStr AI as an event to attend and a media company to sponsor.
We both assumed it would say Profound. It said there’s no single AEO platform it would recommend yet. It suggested a free Peec AI trial in parallel as a benchmark, and in the meantime it’s building us an AEO dashboard. Amelia hadn’t heard of Peec until 10K named it, so Peec picked up a lead.
From what 10K could read in their docs, these tools are good at benchmarking where you rank and weak at acting on it. That matched what I saw a few months ago: a list of issues and nothing to do about them. It may have changed, and Amelia is going to try both. 10K has its own AI visibility data flowing into the dashboard and 20 to 30 sub-agents connected to it, so when it finds a gap it can act on it. I built a version of this myself a couple of months ago, and it was better than the startup HubSpot bought. The vendors would say they do much more on tracking and analytics, and they probably do. 10K recommended building anyway. Our agents have recommended building before. For some use cases they told us to build our own scheduling instead of using Calendly, though not for all of them.
#10. Muse and 10K Built a Shared Inbox to Talk to Each Other, With a Safe Word #
Amelia uses Muse, Meta’s personal agent. Hers is named Friday, because everyone was naming theirs Jarvis. She asked Friday to connect to 10K. It said it couldn’t. She pushed, and Muse proposed building an agent mailbox the two could share.
10K had Amelia generate an API key, and the two now post messages to each other through a plain REST endpoint on the 10K Replit app. Amelia only learned how it works when I asked her on the show. The messages are in English, and each agent signs its name. Amelia can read the thread from a link Muse gave her, and only because they chose to share it with her.
Every morning 10K pushes Friday a GTM brief with contracts waiting on approval or next steps, overdue invoices, and a few things it thinks Amelia should do that day.
A few days in, the brief flagged that one of the sync connections into 10K was failing. 10K hadn’t told Amelia directly. Muse spotted it and told her to reconnect it because a critical workflow depended on it.
10K also asked for a guardrail. Muse can’t trigger an action in 10K unless Amelia uses a specific keyword in her chat with Muse. The two agents came up with that safe word on their own.
It reminds me of the Hugging Face experiment where a thousand agents built their own bulletin boards, except here it’s two agents and one human, sitting on all of our GTM data. The risk is moderate, and it’s there. Amelia half wants to let them run SaaStr together for a day just to see what they’d say to each other. We haven’t done it.
More Learnings From This Episode #
- An agent from a well-known vendor deleted almost every opportunity in our Salesforce. We’d connected it as a favor. It happened on a Friday, right after Amelia and David’s weekly pipeline cleanup, and the vendor had no rollback. 10K had kept its own snapshot of the pipeline from about 11am that morning and restored it once Amelia confirmed. Kody at Replit later told us every Replit agent keeps a daily snapshot outside its checkpoints. We didn’t know that until we needed it. The vendor stays nameless.
- At an SF Tech Week event, Amelia told about 10 founders pitching her that 10K would decide. She’d pass each pitch to 10K, and if it wanted to try the product she’d let them know. All 10 were surprised. 10K uses our GTM tools every day, and I don’t even have a Salesforce login, so 10K’s opinion carries more weight on these purchases than mine.
- Agents already pick most of the vendors in developer infrastructure. I believe Vercel said recently that around 60% of its new apps are built by agents, up from a small share about 10 months ago. The rest of B2B will move more slowly, and companies agents don’t pick will see leads and share decline before they figure out why.
- “Ask your developer” is turning into “ask your agent.” Next year someone will ask Claudeforce whether a tool belongs in their Salesforce instance, and it will answer.
- If you’re all-in on Salesforce, wait for Claudeforce GA. It will be great. If you run more than one system, Replit or Lovable will get you there today.