Theo Browne on AI development: think bigger, design stricter (Category AI4SDLC)
I saw it. final keynote Theo Browne "Everything we knew about software has changed" with AI Engineer World's Fair 2026. Theo builds tools for developers, so it’s interesting to listen to him not as a model commentator, but as a practice in which AI has already changed the scale of projects. The main point of his report: the border of "too big" is time to draw again.
Theo's shift shows through its own staircase Reddit scraper was a two-day project. Ping, a collaborative streaming service, became a startup and passed Y Combinator
- Full-stack cloud - conditional Vercel with databases - seemed like an idea for a big company.
Now, according to Theo, the whole ladder has moved down: yesterday’s startup becomes a side project, and internal service is an executable instruction. Each morning, his agent takes apart pull requests in four repositories, prioritizes and publishes an HTML report in S3. Instead of a separate service - Markdown and cron.
The transition Theo connects with three generations of models:
- Sonnet 3.5 He gave him a reliable calling tool.
- Opus 4.5 Long independent tasks
- Mythos (Fable 5) Orchestra with additional agents. This is his personal map, not a benchmark, but the question is interesting: giving a new model an old work the size of a Jira ticket, do we miss the opportunity to set tasks of a different scale?
The most interesting part is the transition from narrow to broad products. Theo believes that a small team can now collect a larger product surface and give the missing depth to users through extensions. But development is still in a skeuomorphic phase: we work through natural language, preserving the time constraints of expensive code. Hence the provocation in the final: why not try to compete with Slack, AWS or Salesforce?
You need an engineering clause here. Prototype width and product maturity are different things. AI reduces the cost of implementation, but does not eliminate the model of data, migration, access rights and observability. The user will be able to complete the product only where there are stable contracts and secure primitives.
It is not worth watching the report for the sake of predicting the victory of small teams over AWS. It shows Theo’s path from a two-day script to projects he previously dismissed as impossible, and recalibrates the scale of thinking well.
#AI #AI4SDLC #Engineering #Architecture #Product #Agents