pull down to refresh

Model labs are building foundation models. They’re in the R&D business, spending years and billions training the next GPT-whatever before they even think about products.
Agent labs are shipping products today. They take existing frontier models and turn them into goal-directed systems that actually get stuff done.
Describing agent labs, he says
They ship first, optimize later. While model labs are in multi-year R&D cycles, agent labs are shipping products in weeks and iterating based on real user feedback.
They own the full workflow. Model labs see prompts and responses. Agent labs see the entire trace—file changes, tool calls, test results, user approvals. That operational data is their moat.
They’re domain-specific. Instead of trying to build general intelligence, they focus on specific domains where there’s still “lots of work remaining” - the integration work, the domain expertise, the grunt work that Karpathy emphasizes as the real challenge.
They deliver outcomes, not outputs. This is the key insight. You’re not paying for AI tokens—you’re paying for deployed applications, closed tickets, shipped features, or resolved bugs.
Makes me think of Open Agents.
I take the opposite view. This is no enduring moat. When the labs want this market they will take it. I'm assuming this will come in the form of acquisitions, so a couple of these companies may do okay (OpenAI was on the cusp w/ Windsurf); but they may decide to bypass them entirely, as per Claude Code / Codex.
reply
102 sats \ 0 replies \ @optimism 10h
I've felt like this for a longer time too, but isn't it amazing that the actual LLM agent providers are nearly all for software development? It feels like there is zero traction besides a couple of healthcare oriented ones that are geo-fenced to be US-beneficial only.
reply