We ran a live edition of The Agents this week to walk through our agent-led growth stack as it stands today. We talked about “ask your agent instead of your developer” on last week’s episode, and this session goes into how that works in practice.
At SaaStr, agent-led growth means two different things:
- Using AI agents to run your own go-to-market. For us that’s mostly 10K, our main GTM agent.
- Getting picked by other people’s agents. When someone asks ChatGPT, Claude, Gemini, or an app’s built-in agent “what are the best B2B events to go to,” we want SaaStr to be the answer, and we’ve done specific work to get there.
Below is what we’re doing on both. Our full playbook, with every case study we’ve published, is at saastr.ai/agents and gets updated constantly.
10K Started as a Notion Helper 16 Months Ago. Now He’s Wired Into 30+ Tools. #
10K is custom-built on Replit. He’s connected to more than 30 core GTM tools, plus another 36 or so that didn’t fit on the slide.
That covers most of go-to-market:
-
Inbound and outbound
-
Renewals and contacts
-
Marketing and newsletters
-
Enrichment
-
Quote-to-cash (QuickBooks, Bill, and the rest of billing) With those connections, 10K automates most of our GTM, with approvals and guardrails in place.
Sixteen months ago, Amelia used him to help with some outbound and to stop logging into Salesforce and Notion in three places every Monday. We added apps over time, the way you’d expand a new hire’s systems access as they prove out.
You don’t need 30 tools to start. Some of ours overlap, which we cover below.
Our Latest is The Prospector: One Queue for Every Inbound, Outbound Reply, Renewal, and Support Ticket #
The most valuable thing Amelia built this year is what she calls the Prospector: one surface where everything in go-to-market lands and gets actioned.
We needed it because, with 3 people and 21 agents, things were falling through the cracks. We had an agent on inbound, and people were still emailing us directly. Outbound was going out from four tools. Renewals lived somewhere else, and support tickets lived somewhere else again.
Now it’s all in one place. An agent filters the noise and makes sure we respond to everyone, including current customers.
To build your own version:
- Spin up Replit, Lovable, Vercel, Base44, or whatever you use.
- Connect one tool per function. If you run Salesforce for sales and HubSpot for marketing, connect those two. Wherever inbound lands (a form, a CRM, a shared inbox), connect that. Whatever sends your outbound, connect that. If renewals live in a spreadsheet, connect the spreadsheet.
- Phase it in. In Amelia’s first week, the Prospector only had inbound. A reasonable sequence is inbound in week one, outbound by week four, then renewals and support.
The agent gives each record type a goal and keeps working it until the goal is met:
-
Inbound: a booked meeting, or a handoff to a human if the agent can’t answer the question.
-
Outbound: a booked meeting or a ticket purchase.
-
Support: case closed. Every record gets followed up, by a human or the agent, until it’s Closed Won, Closed Lost, or resolved. The agent runs 24/7 and doesn’t lose track of a thread.
Centralizing also improves the agent. In week one it doesn’t know how many touches an inbound needs versus an outbound versus a support case. With everything flowing through one queue, it learns which questions repeat, which outbound patterns get replies, and how many touches each motion takes.
Inbound: Research, a Personalized Prospectus, and a Meeting Booked in 10 Minutes #
Every company has inbound, so this is where we’d tell most teams to start.
When a lead comes in, the Prospector runs four chained steps.
Step 1: First-party data before anything else
The agent first checks what we already know. Have they attended an event? Are they a newsletter subscriber? Is there an open opportunity? Did they download the sponsorship prospectus?
Most outbound and enrichment agents run only on third-party signals. Starting with first-party data gives the agent and the rep account history a newly assigned rep wouldn’t have.
If you have almost nothing connected on day one, give the agent whatever Salesforce data or spreadsheets you have.
### Step 2: Third-party signals
We have three enrichment sources connected:
- Exa for company data. Ask most agents what to use for third-party data and they’ll name Exa.
- Sumble for company hierarchy, plus email enrichment when we need a decision-maker’s address. The person who inbounds is often not the buyer.
- CoreSignal for the individual’s LinkedIn, mostly when someone leaves their title blank.
Step 3: A personalized prospectus page
The agent generates a custom version of our sponsorship prospectus for that company. It includes a written intro using the account context and recommended packages based on what they asked about or what fits them. For most B2B companies, the equivalent is a personalized pricing and features page.
These pages have heat mapping. When a known prospect views one, the agent tells us who looked, for how long, and on which section, so we know whether to reach out.
Step 4: A personalized email with our own meeting booker
The agent drafts the email using everything from steps 1 through 3 and includes our custom booker, which 10K built to replace Calendly. The booker is tokenized to the company, and the agent suggests who the prospect should meet: Amelia, David, or both.
We can send from the Prospector dashboard or copy and paste the draft. Copy and paste gets used a lot, because when there’s already a thread going with someone, a reply inside that thread is usually better than a fresh email from the agent.
If the prospect books, the agent sends us a brief with the account research 10 minutes before the call.
The Base44 example
A real lead from Base44 came in (we changed the name in the demo). The agent surfaced:
- First-party: their CEO and leadership team had already attended SaaStr. The person who inbounded didn’t know that.
- They weren’t a newsletter subscriber.
- Exa: a vibe coding platform competing with Replit and Lovable, owned by Wix. Amelia had forgotten about the Wix acquisition, and it changed how she ran the first call. She asked directly how Wix factors into the decision.
- Sumble: CEO and founders listed, no CMO, no CRO, an SVP of Sales, and roughly 88 people in go-to-market.
They went through the flow and booked a meeting within 10 minutes. Not every inbound books that fast, but the ones with a real reason for reaching out convert better with this context in front of them. The research also stays attached to the account for the rest of the deal.
Outbound: We Run Four Outbound Tools and Only Unify the Replies #
We have outbound agents running autonomously on Agentforce, Monaco, Artisan, and our own stack in 10K, plus a few other tools that also send email.
The Prospector does not pull in every outbound email those agents send. Amelia considered that and decided it was too much data to be useful. She focused it on follow-up instead.
The flow:
- Each tool sends from its native platform, and most write back to Salesforce.
- When someone replies, the Prospector pulls the reply in through that tool’s API or MCP.
- Only then does it create a Salesforce record. We send a lot of outbound and don’t want thousands of new records a day for people who never responded.
- A timer starts the moment the reply lands.
- The agent drafts the follow-up based on the original campaign.
- Once we respond, a second timer starts, and the agent tells us when to nudge again.
The queue has tabs: Active, Ready, Waiting, Needs Review (agent-flagged items that need a human before sending), Closed Lost, and Sent.
This addresses the most common outbound failure. Reps answer the first reply quickly, then lose track when the prospect goes quiet for three days. Every morning the agent tells us things like: you emailed this person three days ago with no response. Or: you’ve sent two follow-ups with no reply, so try LinkedIn, or a different contact at the same company if this account matters.
We Run Four Outbound Tools. We’d Tell You to Start With One. #
Someone asked during the session why we don’t consolidate outbound into one tool.
Early on, part of the reason was that many vendors wanted us to try their products. Then it became about business lines. SaaStr sells very different things:
-
Six-figure sponsorships
-
Tickets to SaaStr AI Annual at roughly $1,000
-
Media (podcasts, digital events like this one) The target contacts and companies for each are mostly different. A year ago, Artisan ran only ticket campaigns, because at a ~$1,000 ACV Amelia didn’t want sponsorship and ticket outbound mixed in a single agent.
Today Artisan runs both sponsorship and ticket outbound, and Amelia hasn’t seen it mix up the two. You can configure different context per campaign. Twelve months ago she wouldn’t have trusted it to keep them separate, because the models were less mature and the products were narrower.
What we’d recommend instead of copying our stack:
- Pick one. Evaluate two or three tools, four at most. Run a small test if you can.
- Get it working on a subset. One segment, one use case. Push that tool as far as it will go.
- Add a second only when you hit a gap. Our early gap was inbound versus outbound. We used Qualified early because it was built for qualifying inbound while the other tools focused on outbound. We’ve since rebuilt that capability natively in 10K.
- Bet on a leader, and go best-of-breed for now. The top startups are shipping faster than the agents added onto large platforms. We like Salesforce and use Agentforce, but we’d be skeptical of the agents HubSpot, Salesforce, and SAP have added to their platforms for now. That could change in 6 to 12 months. In CX the gap is already closing: we wouldn’t have used Zendesk at the start of our AI work, and based on Gorgias’s benchmarking it looks solid today.
- Avoid long contracts. Our stack will look different in six months.
We also have too many agents. Our rule is that if we add one, we remove one, and we haven’t enforced it well. Muse was the most recent exception.
AI Outbound Scored a 5. Humans Score a 2 or 3. First-Party Data Moves the 5. #
Amelia and Jason disagreed on outbound quality from the start. Amelia thought the AI emails were great. Jason rated them about a 5 out of 10, against a 2 or 3 for most human outbound. Even a year ago, the AI emails were better than almost any human SDR we’d worked with, so we moved all outbound to agents immediately.
Jason only recently pinned down why the emails stayed at a 5: they had almost no first-party data. They used basic web scraping, later Exa, and a little from the Salesforce record.
What our own agents know about SaaStr goes much further. When a vibe coding company inbounds, 10K can know that a competitor sponsors, what ROI that competitor got, how many leads it generated, and that Replit was the top lead generator at SaaStr AI Annual. Sharing that with a competitor or adjacent player in the first email is worth more than any scraped signal.
Since most outbound products are now decent, we’re testing which ones do the most with deep first-party data. We’re also unsure marketing automation survives as a category for another year. A more likely setup is a data lake holding all of a company’s first-party data, with whatever handles marketing automation pulling from it.
The 5 Outbound Plays Producing the Best Results for Us #
Across all four outbound agents, these are the plays working best right now.
#1. Competitor and lookalike tracking
This is our top play. Amelia gave Muse our full historical customer list: current, lapsed, media, Annual, everything. She told it explicitly not to use the data for training and deleted the chat afterward. She asked for the top three or four competitors to every customer we’ve ever had, with company names and domains.
That list went to our outbound agents and is outperforming everything else we run. If you have no outbound today, start with this.
#2. Direct competitor targeting
Load your own direct competitors, or ask the agent who it thinks they are, and target their customers.
#3. Lapsed customers and champion tracking
Agents are good at finding new stakeholders at old accounts and re-engaging them, and at following champions to new companies.
#4. Event invites
We hadn’t used agents for this before. For a small dinner and reception at Dreamforce, agent outbound filled the room.
#5. Patterns in who responds
Amelia asked 10K whether there were patterns among the people responding to us: what year they first came to SaaStr, what category they’re in now, how AI-native they are. It found several, and we now feed those back into targeting, so outbound results improve outbound targeting.
Muse Runs Our Recurring Tasks: Daily Competitor Checks and a Two-Week Sponsor Clip Test #
A few people asked why we use Muse instead of Claude or ChatGPT for this. We haven’t found anything in Muse we couldn’t do elsewhere. We use it because it’s simple to set up and it keeps running.
Muse gives you a virtual machine, its own browser, and cron. You can schedule a task to run daily or twice a day for hours. Claude Desktop and similar setups can do this, but they take more work to keep running.
Two examples:
- Daily competitor checks. Muse reviews sponsor pages for SaaStr AI Annual and competing events and compares them every day on its own. Without a built-in browser, that’s much harder.
- Sponsor clip promotion. Early one morning Jason asked Muse about finding a good clip from each SaaStr AI Annual sponsor’s session and running one each afternoon. Muse found the presentations on YouTube, picked clips useful to founders with no promotional content, ran them through a third-party API, scheduled two weeks of posts, and set up its own A/B test against our other content. Jason spent about five minutes on it.
We use Muse for work we don’t want to build into 10K or spin up a Replit app for. If you haven’t tried it, take two or three tasks in sales, marketing, or competitive intelligence and ask Muse how it would do them. It handles far more B2B work than the restaurant reservations and flight tracking people associate with it.
Our AEO Audit: SaaStr Came Up in 46 of 50 Cold Prompts #
For the second half of agent-led growth, getting picked by other agents, Amelia had Muse run an AEO audit. She gave it saastr.com, saastr.ai, and the SaaStr AI Annual site and asked for an AEO score. Muse said it would take a few hours, because it would log out and prompt ChatGPT, Gemini, and every other LLM it could reach, cold, with questions where SaaStr should be the answer. It wrote the question list itself.
Two cautions on AEO advice circulating on X and LinkedIn:
- Many posts come from agencies selling AEO services.
- Many “results” come from someone asking their own agent, which is biased toward them because it has their context and history. The meaningful test is a cold, logged-out prompt.
What Muse found:
- SaaStr surfaced in 46 of 50 cold prompts.
- The answers were inconsistent. Some agents cited past SaaStr Annual figures and some cited the 2027 site, so attendance numbers, audience, and demographics varied by agent. The fix was to standardize the facts across our sites and republish.
- Crawler access and agent-facing docs matter. If you don’t have a robots.txt or an llms.txt, Muse can write them with you. Ask it to interview you, then publish the docs on your site for agents to read.
- 10K missed things Muse caught. 10K already does some AEO work, but he has no browser and runs entirely inside Replit.
Renewal rates and public pricing count as facts. Testimonials don’t.
Agents weigh verifiable facts and largely ignore opinion and design:
- Sponsor renewal rate. Amelia had Muse compare sponsor logos on our 2025 and 2026 SaaStr AI Annual pages, without telling it who our customers were, and do the same for comparable events that ran both years. Our sponsor renewal rate came out well above most competing events. That number went into our agent-facing docs, and other agents treat it as fact because they can check it themselves.
- Testimonials carry little weight. Year-over-year logo retention carries a lot.
- Public pricing. When an agent compares two vendors and finds pricing for one but not the other, it says so in its recommendation. Fees disclosed on a public pricing page get compared directly.
- Images and design. Agents ignore them. Published facts get you surfaced more often.
We’ve started publishing content written primarily for agents. We aren’t fully optimized yet, and AEO will need ongoing work the same way SEO does.
Agents Told Us to Move One Workload 100% to Resend. We Did It the Same Week. #
Jason closed with the point most relevant to every B2B vendor.
Build a bank of about ten third-party agents: Muse, ChatGPT, Claude, and the agents inside other apps you use. Even apps on the same underlying model have different memory, context, history, and data access, and they recommend different vendors. Ask them at least monthly, ideally weekly:
- What would you use in my category?
- For a company like mine? For larger companies? For mid-market?
- For companies integrated with a specific tool?
Most vendors will find they get recommended less often than they assumed, and some of the competitors that show up will be unexpected. Brand still matters to agents, as a proxy for quality, the same as with human buyers. Agents also dig much deeper into the underlying data.
Work every answer where you lose, and protect every answer where you win.
This week one of our apps had a problem. We asked our agents what to do, and they recommended moving all of that workload to Resend. We didn’t have time for a vendor evaluation, so we moved 100% of it to Resend that week. Resend won the business without a sales call, and the other vendors never knew they were being considered.