CharlieKey takeaways from your favorite podcasts
← 13 October 2026Latent Space

Feature requests went from 6 months to minutes after we killed our SaaS

Latent Space ran a contest to replace a $43,000/year event-software SaaS with AI-built clones, finding that agents can build credible replacements but depend on specs, human testing, taste and ongoing ownership.

Watch on YouTube ↗

Charlie's summary

  1. Sean ran Kill My SaaS because buying software means $43,000 a year and 3 to 6 month waits for change requests.

    Sean was quoted $43,000 a year for four events, and uses only about a third of what the vendor bundles, with no discount for the rest. Change requests often take “3 months to 6 months or never.” He had long banned internal software builds to avoid not-invented-here syndrome, but changed his mind because agents are now good enough. 72 teams submitted, judged by his real, non-technical events team.

    ▶ 0:28
  2. Mid-tier business software is hard to clone because of overlapping roles and portals, and three features tripped most builders.

    Sean's example is neither Salesforce nor a one-day vibe-code: speakers, sponsors, invited speakers and attendees overlap, and roles get invented on the fly. Most submissions failed on three things: separate portals for each audience, drag-and-drop scheduling that prevents double-booking, and adding an invited speaker without faking a CFP submission. None were well specified in the docs. You only find them by using the product like a real user.

    ▶ 3:03
  3. An AI judge is good for hill climbing but a poor predictor of what real users prefer.

    Sean had Claude log in as each role and click through 20 scenarios. Of 56 apps, the top seven landed within 2.3 points, which is saturation. But six of the AI judge's top 10 were rated weak by Kelsey's human team, and the winner placed 19th on the AI leaderboard. Sean suggests publishing scenarios so builders can hill climb, then judging with humans. He also says the contest looked for people who “code with taste.”

    ▶ 7:10
  4. Builders mixed models, and the best setups never let a model approve its own work.

    Claude Code and Codex were about tied, and 40 of 66 submissions used both. Sean found three patterns: a smart model plans while a cheaper one writes; one model builds while another checks, usually Codex for browser testing; and no model approves its own work. Green Room HQ ran a Claude team and a Codex team that could not merge until the other verified. Sean notes agents go easier on their own work, as humans do.

    ▶ 10:13

Give your email to unlock the rest of the key points — for every episode.

Do more with Charlie

Free — sign in with your email, no password.