Agent Labs: Welcome to GPT Wrapper Summer (www.latent.space)

“Agent Labs” is a new investment and product thesis arguing that the next wave of high-growth AI companies will be built around shipping specialized, product-first agents rather than chasing SOTA LLMs. The term contrasts with “Model Labs” (the big R&D-first plays) and with “Neolabs” (a catch‑all for novel modeling approaches). Agent Labs (examples: Cursor, Perplexity, Cognition, Sierra, Lovable, Gamma, plus product-forward transitions like Notion, Vercel, Glean, Replit) prioritize real outcomes, outcome-based pricing (hundreds to thousands $/month or per outcome vs. consumer $20/month subs), rapid harness iteration, human-in-the-loop controls and higher margin cashflows. The core business claim: selling end results (task automation, workflows, developer UX) beats competing token-for-token on model capability and price. Technically, Agent Labs treat agents as systems—bundles of model, prompt, memories, tools, planning, orchestration and auth—so competitive advantage shifts from raw model scale to system engineering, orchestration and domain-tailored post-training (RL fine-tuning). Two signals accelerate the thesis: OpenAI’s resource split (inference ~28% of compute) and its public pivot toward an “AI Cloud” serving third‑party apps, plus Anthropic’s massive infra expansion. With pretraining nearing diminishing returns, the era of posttraining/RL and domain-specific continued training favors Agent Labs that can start from strong open weights and close capability gaps fast. The practical implication: expect more startups building vertically integrated agents, continued R&D inside product teams, and a marketplace where model selection is one component of a richer systems stack.