[Editor’s Note: This guest post is by Marcelo Calbucci, a longtime Seattle tech and startup community leader.]
This month, I ran a survey with early-stage founders from Seattle-based Foundations about their use of AI instruments and brokers. There have been some surprises within the knowledge — and never within the path you’d anticipate — and developments which are value speaking about.
The pattern measurement represents 22 startups with one-to-five software program engineers every, for a complete of 42 folks. What makes this cohort helpful to grasp is that they’re AI-native startups, began throughout a time that AI can code. This provides us a glimpse into the way forward for tech corporations.
The primary query I requested on the survey was concerning the share of manufacturing code being written by AI. I wrote this query explicitly to exclude unit checks, scripts, paperwork, and different artifacts that aren’t associated to the core worth proposition of a enterprise. If you understand one factor about AI coding, it’s that it generates giant volumes of unit checks, readme recordsdata, and scripts. None of that pertains to the code that delivers the worth to the client.
Right here’s the stunning reality: out of the 22, 4 startups (18%) mentioned AI is writing 100% of their code. That’s mind-blowing! It doesn’t imply these of us aren’t reviewing and re-prompting the AI to refine the code. Nonetheless, it means they aren’t typing code in an IDE. There are 11 startups (50%) the place AI is writing 80-99% of the code. Including the 4 the place AI writes all the pieces, 68% of startups have AI write over 80% of the manufacturing code. On the opposite facet of the spectrum, three startups (13.6%) mentioned that AI is writing lower than 50% of their code.
Select your weapons
From the information that Cursor will get within the press, you’d suppose utilization for this cohort is near 100%. In our pattern, out of 42 programmers from 22 distinctive startups, “solely” 23 of them (54.7%) use Cursor. On common, Cursor programmers spent $113.63/individual in September. The preferred instrument, although, is Claude Code, with 64.3% of programmers utilizing it and spending $167.41/individual in September. Claude is the popular instrument for startups, with 16 of the 22 (72.7%) utilizing it.
After Claude and Cursor, there’s a huge cliff, with OpenAI Codex coming in a distant third place with seven of the startups utilizing it, representing 12 of the 42 programmers. On common, bills with OpenAI Codex got here in at $48.49/individual in September. The fourth and fifth locations have been GitHub Copilot and Gemini CLI by Google. They’d 9.52% and 4.76% of programmers utilizing it, respectively.
On common, every software program engineer spent $182.55 within the high 5 AI instruments talked about above, with some startups spending over $400/individual.
Founders additionally talked about they use quite a lot of instruments to create manufacturing code, together with Lovable, Devplan, Mentat, Manufacturing facility.ai, Jetbrains Junie, Warp, and Figma.
Roadblocks
When requested about what’s stopping extra use of AI for coding, the primary grievance was the standard of the code. One other hindrance to quicker adoption is the educational curve to get the agent to do what you need.
When it comes to frustration, this group raises three key points. First, the standard of the output, requiring appreciable rework. Second, a mismatch between expectation and actuality based mostly on what everyone seems to be listening to. Lastly, the commonest frustration — and I undoubtedly empathize with this one — is managing the context and coping with giant code bases.
What’s subsequent?
Within the survey, I requested about their intention to proceed utilizing AI instruments and brokers to help with product improvement. The survey requested the founders in the event that they supposed so as to add, take away, enhance, or lower utilization of every instrument. The largest winner, by far, was Codex, with 9 startups (40.9%) saying they aren’t utilizing it but, however plan to make use of it in This autumn. As soon as I normalize the info to account for what the expectations are for This autumn, Claude will keep its management, however Codex will match within the variety of startups. Cursor and GitHub Copilot will development barely decrease, every with one startup saying they’ll cease utilizing it. Lastly, the Gemini CLI may see a small enhance in adoption, with three startups claiming to provide it a attempt in This autumn.
Opposite to the various different facets of software program engineering like selecting a cloud supplier, a language, or database, AI instruments and brokers aren’t a zero-sum market. On this survey, 68.2% of startups used a couple of AI instrument to help in manufacturing code improvement. Based mostly on their acknowledged intention, that quantity will develop to 86.4% in This autumn.
