Agentic AI in Practice: Lessons from SaaStr on Scale, Reliability, and Human Oversight

I watched the Jason Lemkin and Amelia Lerutte presenting their bot net over at SaaStr AI.

First, mad respect for Jason and team for going deep and hard on agentic, working in tandem with different vendors, including Lovable and Replit side by side.

Second, it confirmed our botnet is more advanced and there’s not a lot more we can do at this stage – primarily due to tons of missing gaps in humans managing robots or tasks that literally require manual clicking or dragging files into folders or collecting contacts between different networks, and many many more.

Third, it outlined hallucinations the way we see them too – memory is solid, context is vectorized, things run smoothly – and occasionally break.

Fourth, Jason also outlined how the codebase goes from 16k LoC to 46 LoC and that will bring slop, and larger context = more errors. We see that too. Individual siloed agents are more stable, but bringing that microservice architecture is generally a worse call when they share a database and core skills.

Fifth, Jason and Amelia confirmed that AI emails and outreach beat 95% of the marketers in market, and bots can manage 100+ partners or sponsors easily, while CSMs or accounts want to stick to “5 clients”, potentially missing some escalations (the reason Gong and Fathom exist), among other human limitations.

Agents are here to stay. They can save 2.5 hours of manual research in 6 platforms when you have access to 3. They save 4 syncs between people and waiting for hours between them. And the response takes 90 seconds. Even with “are you sure” and “ask Ed to validate that”, you get 5 minutes and a 3x validated data set.

A lot of that needs a human to manage and expertise to navigate or iterate on. Asking a colorblind analytical guy to do designs is a red flag. Asking a philosopher-painter to run statistical modeling over P&Ls is yolo.

But not using custom agents on your data set heavily in 2026 is stone age-alike.