CharlieKey takeaways from your favorite podcasts
← 17 October 202620VC with Harry Stebbings

Horizontal vs Vertical Agents: Who Wins? | The Future of Inference Market | LangChain CEO

Harrison Chase argues that agents stay hard to build, that value accrues to domain experts running vertical agents on open, controllable harnesses, and that demand for inference and agent tokens is far larger than expected.

Watch on YouTube ↗

Charlie's summary

  1. Agents are hard because non-deterministic models cascade errors, and Harrison says the hardness is underrated

    Harrison says models take unbounded natural-language input and give different outputs for similar inputs. Agents chain many such calls with tools, so errors cascade. Getting this system to behave reliably is “the hardest part.” Progress came from better models and better harnesses around them. In the quick-fire he names this as the thing most people underrate.

    ▶ 1:29
  2. Build by guessing which parts are short-term hacks and which will always exist, because models erase product

    Harrison says LangChain keeps erasing parts of its product as models improve, a principle he credits to Cursor. Early on LangChain steered tool calling by having the model emit JSON and parsing it. When OpenAI shipped native tool calling, the team switched fully. His rule: the model will never know your proprietary data or call APIs itself, so the connecting system stays. Hacks like chain of thought and planning get absorbed into models.

    ▶ 3:32
  3. Own your intelligence: keep model, harness and context open, because memory tied to one lab locks you in

    Harrison argues the state agents accumulate, especially memory, should not be linked to one model provider. Switching models is offensive (use the best new model) and defensive (providers can't raise rates). Labs train harnesses for coding, which is in distribution for them. Outside coding, such as sales, research or law, you want control over the model, the harness and the context. He calls this an open and controllable agent architecture.

    ▶ 6:06
  4. Vertical agents win because domain experts can write the verifiers; start at context and harness, not the model

    If value comes from verification, Harrison thinks the winners are domain experts who can quantify what good looks like, and building evals is very hard. He is “massively bullish on vertical agents.” His practical order: engineer the context, then tweak the harness, and only post-train a model once you have users. Context and harness give much higher ROI. He would not post-train a core model to start, but might in five years with heavy usage.

    ▶ 9:08

Give your email to unlock the rest of the key points — for every episode.

Do more with Charlie

Free — sign in with your email, no password.