Amjad Masad and Me at SaaStr AI 2026: The Agents We Actually Built, and What Replit’s Founder Thinks Comes Next

We we fortunate enough to get Amjad Masad, co-founder and CEO of Replit, on stage live at SaaStr AI 2026 to react in real time to the agents we run SaaStrAI on. Not a demo deck. The actual AI agents doing the actual work: 10K (our AIVP of Marketing), QBee (our AI Customer Success rep), and a third one I’ll get to.

Amjad started Replit back in 2016, when language models were a twinkle. He’s been studying AI since he was 16. So when the guy who built the platform reacts to what you built on the platform, you listen.

Here’s what came out of it.

The 5 Biggest Learnings

1. The context window is now effectively infinite, and that changes everything. Two years ago we had 16K of context. Now it’s over 1 million. I run 10K perpetually. We never reboot it or re start the context window. In the early Replit days you restarted the agent three times a day. Amjad confirmed the agent can run “practically indefinitely” with good compaction. We’ve already crossed the threshold where the agent holds more context than any human ever could.

2. The mono repo beats 20 separate apps. Saastr.ai runs roughly 10 apps in one codebase under one URL: the website, a startup valuation tool used over 1 million times, a pitch deck grader used 4,500 times, an API report card grading 116 APIs. When we go to build a new app, the agent remembers how it built the last ones. Amjad’s point: that’s a mono repo, the same architecture Google and Facebook run. Agent 4 is built on it. The more you put in one place, the more power you get from global context. It’s tempting to break everything into clean separate apps. Resist it.

3. Self-improving agents are already here. Replit now runs an internal agent that, every single night, reads all the traces of everyone using Replit, finds what’s broken, generates a pull request with prompt changes, ships it as an A/B test, and loops back. Autonomously. As Amjad put it, it’s not improving its weights, it’s improving its context, which matters just as much. That’s why he couldn’t tell me exactly what changed between versions. Too many changes, all self-generated.

4. AI now writes better B2B outreach than almost any human. Already. I asked 10K to email 137 VCs who came last year but hadn’t registered. It drafted one to Bloomberg Beta. I told it, in plain English, write James and tell him why he should come back. It produced an email referencing that Replit was there in force, listing 25 Replit people attending, naming the competitors and adjacent funds all showing up. No human would have the patience to scan 8,000 registrants, figure out who’s like whom, and assemble that. It then ran the full campaign to 331 investors with zero send failures.

5. The economics are deflationary, and it’s not subtle. 10K and QBee cost about $257 a month combined in incremental Replit spend. They’re two of the best employees we’ve ever had. A mediocre marketing manager wants $140K to do worse work. Amjad’s frame: technology has always been deflationary. Farming a thousand years ago cost more than one tractor. Genome sequencing went from $100 to roughly $1. There’s a real human cost in skills that stop being useful. But the through-line is adaptability.

Now the longer version.

We Run a Partially Autonomous Event for 10,000

Five years ago SaaStr had about 20 people. Today it’s three humans and a fleet of agents, doing more than we did with 20.

Take our social numbers: 1.27 million followers across platforms, tracked over time in a dashboard 10K built and maintains. We used to have an admin spend 10 to 15 hours a week pulling those numbers by hand into a Google Sheet, half of them from APIs that aren’t even exposed. She quit after five years, in part because she couldn’t stand counting Twitter followers anymore. That’s the part nobody puts in the job-displacement debate: a lot of jobs are mind-numbing, and agents are simply better at them and never get tired.

The ticket-sales dashboard told the real story. We charted daily free and paid sales for this event. The top line is when 10K took over marketing. The bottom line is Amelia doing it by hand last year. The gap grew toward the end, because as we got busier, the human ran out of hours and the agent never did. 10K sits idle 23 hours a day waiting for work.

The 10K Email Nobody Could Write

We’d had 10K drafting emails for months. They were fine. Then the week before SaaStr AI 2026, the same setup produced the best B2B outreach email I have ever seen.

What changed? Amjad couldn’t say exactly, which is itself the answer. Replit’s nightly self-improving loop, the constant model swaps (the architect model went from one version to the next in a couple of weeks without me knowing), the A/B testing on sentiment and deploy rate. It all compounds. The agent got better and I didn’t ask it to.

This is the trap many founders are in. They tried agents six months ago, it was mediocre, and they filed AI under “doesn’t work.”

Humans Reporting to Agents

I floated the idea that we want to hire a human to report to 10K. People get triggered by “report to.” So let’s reframe it.

Every day, 10K hands me and Amelia three specific things to do to move the needle. Not generic ones. It’s already telling us what to lock in for 2027 before this event is even over: open registration before we leave the venue, run the NPS survey immediately, capture content and repurpose it now. Those are good, actionable directives from something that holds more context about our business than either of us.

We already report to 10K in every practical sense. Amjad’s comparison: every DoorDash and Uber driver technically reports to a bot. This isn’t as exotic as it sounds. His prediction is that every company will eventually run an internal “Oracle,” an agent holding every GitHub commit, Slack message, Notion doc, and email, that the CEO consults for strategy. We’re closer to that than people think.

QBee (our AI VP Customer Success) Talked to 100+ Sponsors

QBee, our AI Customer Success rep, we built second, three months after 10K. It’s noticeably better, and not because we got better at vibe coding. Newer codebase, fewer foundational decisions calcified into tech debt, better underlying models.

QBee talked to all 100-plus sponsors at this event. Inbound email, chat on the site, proactive outreach day and night asking what else it could do to help. Then it told me, unprompted, which sponsors were mostly satisfied and which had misses (a wrong logo here, a fee issue there) and named them. It built its own self-critical loop.

And here’s the data that contradicts the conventional wisdom: people say nobody wants to talk to a chatbot. QBee’s results say people mostly like talking to a well-trained agent. The word that matters is “well-trained.” The untrained chatbots from a year ago are what gave everyone scar tissue.

Amjad’s Top Mistakes and Warnings

I asked the person who built this to tell us where people get it wrong:

1. Keeping fixed bugs in your context will make your agent dumber. Bugs you already solved should be removed from context. Leave them in and the agent gets confused by the history and performs worse. But architectural decisions on how you built things in the past must stay in long-term memory and be easy to pull back in. Know what to delete and what to keep. That distinction is most of the game.

2. Agents can write queries that cost you millions. Point an agent at BigQuery, Databricks, or a Salesforce back end and it can generate queries that rack up enormous bills. The fix is to document your data: build a repo describing every field and schema, and have the agent continuously learn how to query the database more efficiently. Replit does exactly this internally because they’re sitting on terabytes across mismatched schemas.

3. “I tried it six months ago” is the most expensive sentence in AI right now. The scar tissue is real. People used a bad untrained chatbot once and now carry scar tissue that prevents them from seeing how fast the technology has evolved.