Overview of 20VC with Harrison Chase (LangChain)
This episode is a deep dive into the current and future state of agents with Harrison Chase, founder of LangChain. The conversation covers why agents are still hard to build, where value really accrues in the stack, how model routing and memory should work, why vertical agents may win over horizontal ones in many domains, and why inference demand is much larger than most people expected. Chase also shares practical advice on building AI products, managing coding costs, fundraising, and why founders should consider taking a discount to work with tier-one firms like Benchmark and Sequoia.
Key Themes and Takeaways
Agents are still extremely hard to build
- The core challenge is non-determinism: LLMs are smart but unreliable, and small prompt changes can produce very different outputs.
- Agent systems compound that problem because they are multi-step workflows with tool calls, memory, routing, and other sub-agents.
- LangChain’s value is in helping developers build the harness around the model so agents behave reliably.
The stack: context, harness, model
- Chase frames “owning your intelligence” as controlling three layers:
- Context: data, memory, APIs, tools
- Harness: orchestration, routing, workflows, logic
- Model: the underlying LLM and eventual post-training
- The cheapest way to start is to control context.
- Next comes harness engineering.
- The highest ceiling, but hardest layer, is model control.
Model progress keeps deleting product surface area
- As models improve, companies need to erase or replace parts of their product that become unnecessary.
- Founders should distinguish between:
- short-term hacks that will disappear
- durable abstractions that will remain necessary even as models improve
- Example: early tool calling relied on JSON parsing; native tool calling made that workaround obsolete.
Vertical agents likely beat horizontal ones in many real-world use cases
- Chase is strongly bullish on vertical agents because domain experts are best positioned to define what “good” means.
- The people who can write the best evals/verifiers for a domain will capture a lot of value.
- For vertical products, he recommends starting with:
- Harness/context engineering
- Then post-training only once there is real usage and data
Routing should usually live in the harness, not the gateway
- For domain-specific agents, routing needs awareness of state, memory, prompt cache, and task context.
- That means routing is generally better at the harness layer.
- Gateway-level routing can still make sense for:
- general chat
- coding
- situations where you cannot control the harness
- His view: routing is not a commodity, but it is highly domain-specific and harder than many assume.
Evals and traces matter, but aren’t enough on their own
- Chase likes internal, domain-specific evals, but distrusts generic public benchmarks.
- Traces are valuable because they let teams:
- identify failure modes
- update memory/context
- improve harness code
- post-train models later
- The key is to feed insights back into the system; traces alone do not create value.
Memory is likely the next major agent layer
- Chase thinks memory will be ubiquitous in a few years, but it is still highly unsolved.
- The hardest part of memory is not the database — it’s deciding what should be remembered.
- Current memory approaches are too general; he believes the best solutions will be domain-specific.
- His team tried a commercial memory product and shut it down after failing to find product-market fit.
Market and Business Implications
Inference demand is much bigger than expected
- Chase says he underestimated demand for inference and coding agents.
- The market is so large that “everything just works” — there is enough demand for multiple winners.
- This is why he sees strong opportunities across:
- inference providers
- infra primitives
- coding tools
- agent platforms
Costs are rising even as models get cheaper
- Frontier model pricing has improved, but usage is exploding, so total spend still rises.
- In coding agents especially, token consumption can get out of control quickly.
- His recommended operating pattern:
- Visibility into spend
- Basic cost controls
- Harness improvements to optimize cost and efficiency
Accuracy still matters most
- In developer surveys, the top priority remains accuracy, followed by latency and cost.
- For enterprises, security matters too, but still usually below accuracy and cost.
Open models are useful, but mostly for narrow tasks today
- He thinks open models and post-trained models are increasingly useful, but mainly for:
- routing
- guardrails
- extraction
- summarization
- They are not yet the default core engine for most leading agentic loops.
- He expects US open-source models to catch up with strong Chinese open models over the next year.
Where he would and would not invest
Would invest in:
- Inference infrastructure
- Lower-level infra primitives like:
- Temporal
- Modal
- Supabase
- Convex
- Vertical agents
- Strong technical teams with clear domain expertise
Would avoid or be cautious on:
- Customer support, which he sees as overcrowded
- Humanoid robotics, which he believes is far ahead of current capability
Company Building and Fundraising Advice
Tier-one investors matter
- Chase strongly believes it is worth taking a discount for Benchmark, Sequoia, and similar top-tier firms.
- His experience: these firms act like true partners, not just capital providers.
What top firms actually do
- Benchmark: more focused on core company building, positioning, and being a sounding board
- Sequoia: broader operational help, especially recruiting, marketing, and network access
- He credits both with being meaningful long-term partners in LangChain’s growth.
Fundraising is a game, but relationships are better
- He dislikes performative fundraising tactics and prefers straightforward relationships.
- The best rounds, in his view, come from years-long relationships and mutual trust rather than posture.
Memorable Opinions and Predictions
- Model routing is important, but harness-aware routing is the future.
- Memory will be everywhere, but only after better domain-specific solutions emerge.
- Vertical agents will be a major category, especially where domain knowledge can be encoded into context, harness, and eventually weights.
- Agent-to-agent interactions may become the norm for consumer and marketplace experiences.
- Creative AI may be the most exciting next frontier, beyond enterprise efficiency gains.
Notable Quick Hits
- Biggest change of mind: the inference market is massive
- Hardest technical build: their own trace database, SmithDB
- Single favorite VC meeting: first pitch with Benchmark
- Best long-term advice: build fast, stay technical, and don’t underestimate how hard agents are