20VC: Horizontal vs Vertical Agents: Who Wins | Model Routing: Commodity or Increasing in Value | The Future of the Inference Market | Why Founders Should Take Discounts for Tier 1 Firms with Harrison Chase, LangChain

Why the agent harness, not the model, is where value lives

Agents are only as good as the harness and context around the model, and that layer is where durable value sits. Models are nondeterministic and keep changing, so teams need control over their harness, domain-specific evals and routing that knows their task. Chase argues this favors vertical agents built by domain experts, running on open and neutral infrastructure.

Takeaways

• Neutral harnesses matter for owning your intelligence. Agent memory and context should not be locked to one model provider. Chase says owning your intelligence means controlling the model, the harness and the context, which is why he wants an open, controllable harness rather than a provider's closed one.
• Vertical agents should start at harness and context. Chase is bullish on vertical agents, with value going to domain experts who can write verifiers and evals. He would not post-train his own model to start. Harvey and Ramp post-train open-weight models, which suits simple tasks like classification and summarization.
• Smart routing belongs in the harness, not the gateway. The harness holds more state, such as prompt cache state, so routing must be harness and domain aware. Chase says OpenRouter usage mostly reflects model access. Gateways mainly ease porting to harnesses a team does not control, and routing is not a commodity.
• Trust internal evals and traces over benchmarks. Chase distrusts academic benchmarks because models are tuned against them. He relies on domain-specific internal evals plus reading traces. He defines continual learning as a system that improves as it runs, and calls memory the hardest part of agents.
• Coding agent costs are large and need managing. LangChain spends roughly 30% of engineering salary on coding tools, a share that has fallen while token use rose. One plug-in cost about $40,000 in a week. Its fix is visibility, cost controls and harness improvements.
• Inference demand is large enough for many winners. Chase says demand far exceeds his expectations, so Fireworks, Amazon Web Services and Crusoe will all win. He would invest in inference, infrastructure like Modal, Temporal and Supabase, and vertical agents. Stebbings would avoid customer support and robotics.
• Founders should accept discounts for top-tier investors. Benchmark invested when LangChain had two founders, and Sequoia followed two to three months later. Chase says Benchmark helps with the core of the company and Sequoia with recruiting and marketing. He thinks the brand is worth a lower price.

Where the speakers differ

• Gateway routing. Chase prefers routing in the harness. Stebbings pushed for the bull case for gateways, and Chase conceded they help port routing to harnesses a team does not control.