20VC20VC: 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
17 Oct 2026 · 62 min · Episode page ↗The future of AI agents hinges less on raw model intelligence and more on sophisticated 'harnesses' that manage complex interactions, memory, and domain-specific knowledge, making agent development surprisingly difficult and inherently vertical-specific.
Summary
This conversation with Harrison Chase, CEO of LangChain, delves into the technical and strategic challenges of building and scaling AI agents. It challenges common assumptions about AI's ease of use and the commoditization of infrastructure, highlighting the ongoing difficulty of achieving reliable agent behavior, the critical role of data portability, and the unexpected surge in inference costs as agents become more prevalent.
Interesting ideas
Agent Development is Still Hard Due to Non-Deterministic Models. Building reliable AI agents is challenging because LLMs are non-deterministic 'black boxes' that can produce varied outputs from similar inputs, leading to cascading errors in multi-model systems. This means the core difficulty isn't just model intelligence, but ensuring system robustness, making tools like LangChain that 'harness' models for predictable behavior crucial.
The Strategic Value of Neutral Agent Harnesses for Portability. Neutral agent harnesses, like LangChain's, are essential for avoiding vendor lock-in. By separating agent logic and memory from specific model providers, businesses gain the offensive advantage of adopting the best new models and the defensive benefit of rate flexibility. This portability of 'context' (data and memory) allows for true ownership of an agent's intelligence.
Model Routing Will Be Vertical-Specific and Harness-Aware. Intelligent model routing, which directs queries to the most appropriate model, will primarily reside at the 'harness' layer rather than the 'gateway' layer. This is because effective routing requires deep contextual and domain-specific knowledge that only the harness possesses, making it a non-commodity for vertical agents and limiting gateway routing to general or coding tasks.
AI Memory Remains Largely Unsolved and Domain-Specific. Despite its obvious importance, robust AI memory is an unsolved problem, largely due to the difficulty of getting LLMs to effectively perform 'meta-reflection' and distill observations into valuable, durable information. Memory solutions will likely need to be highly domain-specific, moving beyond simple databases to agent-driven systems for learning and retention.
AI Costs are Exploding Despite Cheaper Tokens. While individual AI tokens are becoming cheaper, overall AI costs are skyrocketing due to immense demand, particularly from agent-to-agent interactions and multi-agent systems in areas like coding. Enterprises are prioritizing accuracy, but cost control is rapidly rising as a critical concern, necessitating better visibility, budgets, and harness-level optimizations like prompt caching and dynamic tool management.
Questions worth thinking about
If vertical agents become dominant, how will general-purpose AI labs adapt their business models, and what novel coordination mechanisms will emerge between horizontal and vertical agents?
Given the rising costs of agent usage, will the economic incentives for developing new AI applications shift towards 'low-token' or 'efficient-token' use cases, and how will that impact innovation?
If agent memory remains domain-specific, what does this imply for the potential of true AGI or broadly intelligent systems that can learn and generalize across diverse contexts?
Written by AI from the episode's full transcript. It can get things wrong: if something here looks off, tell us at hello@podpings.com.