Why Agent Harnesses Beat Smart Routing — Harrison Chase on Building Reliable Agents
Key takeaways from a 20VC interview with LangChain co-founder and CEO Harrison ChaseDealroom has a profile for this one. Try Dealroom → (October 2026):
Why agents are so hard to build. Models are non-deterministic black boxes with a near-infinite input space, and agents are chains of many model calls touching external tools, so small input changes cascade into unpredictable answers. Reliably harnessing those models is the central engineering problem.
Better models mean erasing the product. As models improve, LangChain has to delete large parts of its product; Chase borrows Cursor's "delete the product" principle. When OpenAI shipped native tool calling, LangChain quickly abandoned its earlier JSON-generation-and-parsing hack in favour of the more native path.
A neutral harness defends against lock-in. Context and memory must stay portable across providers — switching models is both offensive (take the best model whoever ships it) and defensive (a vendor cannot raise rates on you). "Owning your intelligence" means controlling the model, the harness and the context.
Verification is where value lands. If value is derived from verification, the winners are domain experts who can write evals and capture that knowledge. Chase is "massively bullish" on vertical agents and advises founders to start at the context and harness layer, moving to model post-training only once they have users.
Most "home-grown" models are post-trained. Claims that companies are building their own models are largely "BS" — firms such as Harvey and Ramp post-train open-weight models for their domains, and only a few players train from scratch. Post-training currently pays off for simpler tasks inside a harness (classification, extraction, summarisation), while the model driving the core agent loop remains an open question.
Model routing belongs in the harness, not the gateway. A harness holds more state — prompt caching and prompting strategy — than a generic API gateway. Most OpenRouter and Vercel use is single-key gateway access rather than "smart routing", which happens far less than people think and mostly inside harnesses such as Factory's.
Coding tools eat the engineering budget. LangChain spends roughly 30% of its engineering salary on coding agents and tools, yet over the last two months token usage has climbed while the bill has come down.
Inference demand is underestimated. Chase calls the inference market "so big" and admits he underestimated demand; agents could one day generate a thousand times more requests than humans, and agent cost is becoming a real problem — which is why the memory and evaluation layers still feel wide open.
Read more: Watch on YouTube · 20VC
