AI Makes Horizontal Software Worth Trying Again
maps directly to joelclaw's bet that agent loops make broader, weirder platform experiments cheap enough to actually try
Theo Browne frames AI-assisted software building as a shift on the same order as AWS and the cloud: when the core cost drops, the shape of reasonable software changes. The old startup playbook was to pick one vertical, go deep, and avoid competing with horizontal monsters like Salesforce or AWS. That made sense when code, teams, and deployment were expensive.
The clever bit is the inversion: if AI coding tools make the first pass cheap enough, maybe the shitty horizontal product becomes viable again. Not polished everywhere. Not enterprise-complete. Just broad enough that users or agents can go deeper where they need to. Theo uses LakeBed as the example: a deliberately rough cloud for rough apps, with auth, database, deploys, inference, previews, and runtime assumptions baked together instead of glued one dashboard at a time.
That matters for joelclaw because the system already lives in this space: agent loops, evented workflows, local infra, and weird personal-platform experiments that would be dumb to staff like normal software. If the wall is farther away now, the useful move is not just automating the old work. It’s pushing until the system actually breaks, then using the break as design feedback.
The sponsor aside is also worth noticing: WorkOS is pitching agent-native signup through AuthKit, with references to Cloudflare and Firecrawl. That’s the same broader signal: services are starting to treat agents as first-class operators, not flaky browser passengers.
Key Ideas
- AI-assisted coding lowers the cost of code enough that product strategy can change, not just developer productivity.
- Theo Browne compares the current shift to the cloud: cheap infrastructure created new software categories, and cheap code may do the same.
- The old vertical strategy, shown through Vercel versus AWS, was to go deep in one slice because going broad was too expensive.
- The Salesforce example shows why horizontal products are hard: most customers need the common core, but each customer also has a few weird must-have features.
- LakeBed is Theo’s test case for a broad, rough platform where the code itself carries deployment intent.
- Agent-native signup points at a future where services expose flows for agents directly instead of forcing them through human dashboards.
- For joelclaw, the useful pattern is to use agent loops to explore bigger system shapes, not just speed up the same old tickets.