Ad Narrator | 20VC Summary

1. Key Themes

Value in agents accrues to whoever owns the harness, context, and verification, not just the model

Harrison frames agent control as three layers: context (data, memory, tools), harness, and model. Each is a place to build proprietary advantage, in order of difficulty. He argues the model providers' harnesses are trained for coding, so anything out of distribution needs an open, controllable harness. On why harnesses should be neutral: "Being able to change models is very offensive and defensive. So it's offensive in the fact that you want to be able to take advantage of the best model when it comes out no matter who it's coming from. It's defensive in the fact that you don't want to get locked in so they can't raise rates, things like that." 00:09:40 He defines "owning your intelligence" this way: "model, the harness itself, and the context make up kind of like this open and controllable agent architecture, and we think that's a big part of what it means to own one's intelligence." 00:11:26

Vertical agents are the big bet, because domain experts write the verifiers

When Harry asks who wins if value comes from verification, Harrison says: "I think the people who win are the domain experts, the people who are able to kind of like quantify what that value means, which is very hard. Like, building evals is a very hard job... I'm, I'm massively bullish on vertical agents for different domains for, for this reason." 00:12:02 His practical advice: "if you're getting started building a vertical agent, I would highly suggest you start at the harness layer. You do context engineering, you kind of like tweak the harness, and then when you've got users, then you start tweaking the model and post-training. I think it's, it's much higher ROI to start at the context or harness layer first." 00:12:36

A ladder of "owning your intelligence": context → harness → post-train → train from scratch

Harrison lays out increasing cost and control. "The cheapest way is you just control the context. You use an off-the-shelf harness, off-the-shelf model. The next level down is you start tweaking the harness... The next level down is, is you post-train an open-weight model... And then the very kind of like bottom layer is they actually train a model from scratch." 00:13:49 He adds that "if you're training a model from scratch, you're probably a neo-lab." 00:14:01 Companies like Ramp and Harvey "have gotten all the juice they can out of the context, all the juice they can out of the harness, and now they've arrived at basically post-training of these models." 00:14:21

Post-training works today for narrow sub-tasks, while the core agent loop remains an open question

"A lot of the places where we see post-training being efficient now are for simpler tasks like classification, or extraction, or summarization, that might be a tool the agent uses, or a guardrail." 00:15:15 The unresolved question is "for this model that's driving that loop, will those be kind of like post-trained models or will those be kind of like frontier models?" 00:15:34 On open models generally: "they're for more narrow tasks than people think... I don't think they're really driving the the core agentic loops for a lot of these main agents." 00:45:19

Model routing belongs in the harness for vertical use cases; gateways win for horizontal ones

Harrison distinguishes a router from a gateway: "in the harness level, it's in kind of like the code of how it runs, and in the gateway it's just an API call." 00:16:47 Why harness: "in the harness, you just have more knowledge of what's going on. You have more awareness of the state... what's going on with the prompt cache. You want to try to keep the prompt cache warm and available. You can also change the prompt more easily." 00:17:06 The gateway case holds for coding or general chat: "if you say, 'Hi, how are you doing?' you don't need like Opus for that." 00:20:24 Net: "If you're building a vertical agent, I think routing needs to be in the harness. If you're thinking about a general purpose horizontal task like just general chat or coding, it could maybe be in the gateway." 00:21:09 He also says "smart routing... is actually being done like way less than people think." 00:18:26

Memory and continual learning are the biggest unsolved layer, and the hard part isn't the database

"The hardest part about memory is actually not the database at all... The hard part is taking all of the observations that happen... and distilling that into information that's valuable enough to go into this context layer." 00:27:09 He expects the context layer to "look more similar to like an agent around a database than just a database." 00:26:53 He defines continual learning as "a system that gets better over time as it runs," adding an "outer loop" of "take traces, get learnings, fix something, measure it to make sure it improved, and then repeat." 00:25:28 Meta-reflection remains hard because "I think this type of like meta-reflection, I think is something that LLMs actually still struggle at a lot." 00:27:36

Coding-agent spend is the key variable for AI economics, and it is controllable

LangChain spends "roughly 30% of our engineering salary on coding tools. It's actually gone down the last two months, while tokens... has actually gone up." 00:00:03 Their three-step playbook is visibility, cost controls, then harness optimization. Harry argues this ratio underpins frontier-lab valuations: "If that number goes to 5%, then the $2 trillion IPO might be a little bit overpriced." 00:42:30 Harrison agrees, noting "the only place we've seen agents really have effect is in software engineering so far... undoubtedly, for the other domains, it will be a lot more than it is today." 00:42:54

Inference demand surprised everyone: "everything just works"

"Two things that I think I really underestimated from the start of LangChain to now. One is the demand for inference, and two is the demand for like coding agents." 00:46:38 Quoting Eric Vishria: "'Everything just works.' ... I think any inference will work... Fireworks will work. AWS will work. Crusoe will work. I just think there's so much demand for it." 00:47:05 On changing his mind: "I remember talking with the board about whether we should do a gateway a year and a half ago, and we didn't see what value we could add over something like OpenRouter. And I think that was 100% the wrong decision." 00:57:07

The agent-to-agent future: every app ships its own vertical agent

Harrison's consumer thesis is that "you'll have both horizontal and vertical agents, and the horizontal agents will interact with the vertical agents." 00:32:35 Reasoning: apps hold knowledge about you, and "a lot of these apps have a ton of APIs... It's a better experience if there is an agent that kind of like wraps those APIs and exposes a single... natural language interface." 00:32:46 Providers will monetize inside agent responses: "you could do advertising for specific things on on Resy or Etsy or whatever, and include that in the agent response." 00:35:13

2. Contrarian Perspectives

Smart routing is not a commodity, and it won't live in the gateway

The prevailing view is that routing is a thin layer that gets commoditized or absorbed by gateways like OpenRouter. Harrison says it's hard and domain-specific: "I don't think routing is a commodity. I do think people will build it, but I think building it is quite hard, even today. And I think part of the reason for that is the models just change so rapidly." 00:21:24 He adds that a new frontier model can erase the routing value: "a new Opus, or a new OpenAI model comes out, and it's just better at everything. And so then the value of routing kind of goes down." 00:21:24 And on OpenRouter's usage: "a lot of it just comes from being a gateway, not a router." 00:18:05

Your agent product should be partly deleted every few months

Harrison argues a framework business must shrink its own surface area as models improve: "The main way our business changes as the models get better is we basically have to erase big parts of our product." 00:06:29 He cites Cursor's operating principle ("Erase the product") and the example of JSON-DSL tool calling, which LangChain abandoned once OpenAI shipped native tool calling. 00:07:44

Don't post-train at the start; most "own-model" claims are fine-tuned open weights

Many assume vertical leaders must build proprietary models early. Harrison says: "Not to start. Absolutely not to start. But like if we're talking like 5 years in the future assuming that I'm getting a ton of usage, I think it could be a realistic possibility." 00:16:14 He confirms the skepticism that "a lot of it is they are post-training open-weight models, absolutely. I still think that's a valuable thing to do." 00:13:35

Chinese open-weight model fear is mostly unfounded, and US open models will catch up within 12 months

"I don't really think that fear is justified... if you're using Fireworks or hosting it yourself, I think that's just ridiculous." 00:44:04 He predicts "in the next 12 months, we will absolutely have comparable US open-source models," pointing to Nvidia's ecosystem work and Reflection's recent launch. 00:44:48

Pay the discount for tier-one VCs, and being first is more valuable than ever

"I think absolutely it's worth taking a discard for Sequoia, and Benchmark, and and probably other of these top-one tier or tier-one VC firms." 00:00:00 On first-mover advantage: "Anthropic and Claude Code, that wow, what what an advantage that was." 00:51:42 But first is not sufficient: "AutoGPT... was the fastest-growing GitHub project of all time. I haven't heard much from them since." 00:52:04

3. Companies Identified

Fireworks AI

Inference and specialized-intelligence platform for open models; Harry's fund is an investor. Harrison cites it as a core example of the inference boom and of safe hosting for Chinese open-weight models. "Lin and Fireworks, right? They wrote PyTorch. Like, how cool is that, right?" 00:49:47 And: "Fireworks will work. AWS will work. Crusoe will work." 00:47:05

Cursor

AI coding IDE company. Cited for its operating principle of deleting its own product as models improve, and as an example of a brand caught in "lightning in a bottle." Harrison: "Cursor said this really well as well... 'Erase the product,' or 'Delete the product.'" 00:06:29 And: "Cursor, great example. They went very viral, execute against that." 00:50:40

Harvey

Legal AI company. Named as a leading vertical agent that is post-training models. "I think Harvey recently in the past month or so announced something exactly in that vein." 00:15:49 Also: "the Ramps and the Harveys of the world, they have a ton of talent, they have a ton of data." 00:14:21

Ramp

Fintech company with vertical agents. Cited alongside Harvey for pushing context, harness, then post-training. 00:14:21

Factory

Coding-agent company. Cited for the best public thinking on harness-level routing. "I think Factory does a tremendous job about this. They've written a number of technical blogs that are very well done on terms of how they do uh routing." 00:18:26 Also: "Factory and Cognition have also written things about this." 00:17:06

Cognition

AI software-engineering company. Cited with Factory as writing about harness-level routing. 00:17:06

OpenRouter

Model gateway. "OpenRouter, incredible company, incredible usage." 00:18:00 Harrison notes its usage likely comes from gateway access rather than smart routing, and that LangChain's decision not to build a gateway was a mistake. 00:57:07

Vercel

Also named as a gateway: "a lot of why people use OpenRouter and Vercel and other gateways is largely for the model access." 00:18:05

ClickHouse

Database company, "doing incredibly well recently." Harrison says it's still a traditional database pre-dating AI, and that LangChain is both a customer and a competitor with its SmithDB. 00:26:24 "I think there are 500 different companies that we are frenemies with." 00:57:33

MongoDB and Elastic

Named as existing databases great for software-system data feeding a context layer. 00:26:53

Modal

Low-level infrastructure with sandboxes. "Modal, incredible company." 00:48:40 And: "Modal has an incredibly strong kind of like talent team." 00:49:47

Temporal

Low-level infrastructure primitive. "Temporal, incredible company." 00:48:40

Supabase

Backend infrastructure, named as a layer to invest in as more software gets written. 00:48:40

Convex

Backend infrastructure, named alongside Supabase. 00:48:40

Anthropic

Cited for Claude Code's first-mover advantage and as dependent on enterprise agent spend. "Being first, Anthropic and Claude Code, that wow, what what an advantage that was." 00:51:42 Harrison would pick OpenAI over Anthropic at the stated IPO valuations. 00:57:46

OpenAI

Cited for introducing native tool calling and for catching up on coding. "OpenAI's done an incredible job of catching up with with Codex." 00:51:42 On IPO choice: "OpenAI. I think they have more diverse different paths to really tremendous exits valuations." 00:57:46

Claude Code

Anthropic's coding harness. Used as the example of a beloved harness you can only reroute at the gateway layer. 00:19:20

Instinct (the personal assistant; transcribed variously)

A personal-assistant agent. Harrison says it has "nailed" personalization, proactiveness and memory, and it shows brand-driven advantage in a crowded category. "I think Instinct and Muse have done a great job at this and so you see how it makes a difference there." 00:30:32 Also noted in the Resy anecdote about agent polling. 00:34:13

Muse

Personal-assistant agent cited with Instinct for memory and personalization. 00:30:09

Resy

Reservations platform, cited for blocking a user whose agent was polling it. 00:34:13

Spotify and Strava

Used as examples of apps with distinct domain-specific user knowledge that would justify their own vertical agents. 00:33:05

Etsy

Used as an example of a marketplace that could place advertising inside agent responses. 00:35:13

Open SWE (LangChain)

LangChain's in-house cloud-based coding agent, built to control costs with prompt caching and routing. "It's actually the only way I code anymore because I'm in Slack all the time, and so I just message it from Slack, and it kicks off." 00:39:16

LangSmith and Engine (LangChain)

LangSmith is LangChain's observability and evals product. Engine is "an agent in LangSmith that runs over traces and tries to update code." 00:24:11

SmithDB (LangChain)

LangChain's in-house database for agent traces, named as the hardest technical thing they've built. 00:57:23

Deep Agents (LangChain)

LangChain's agent harness. "We build Deep Agents. Deep Agents is an agent harness." 00:31:53

Nvidia

Cited as caring "tremendously" about the open US model ecosystem. 00:44:48

Reflection

US open model lab: "Reflection had a model launch recently. I think it was still a little bit behind some of the Chinese models on some benchmarks." 00:44:48

Crusoe, Nebius, Baseten, AWS

Named by Harry and Harrison as inference providers; Harrison says demand is large enough that "Fireworks will work. AWS will work. Crusoe will work." 00:47:05

Benchmark

Lead early investor in LangChain. "We really thought of them as kind of like a a a third co-founder." 00:53:14

Sequoia

Joined after Benchmark. "Sequoia's style, I think, is they help with like everything. Like, they are just a machine." 00:54:19 Candidates: "we've gotten a lot of our candidates from from Sequoia references." 00:54:30

IVP

Led the Series B via Tom. 00:56:00

Atlassian

Referenced via CEO Mike Cannon-Brookes' skepticism of "own model" claims. 00:13:13

Suno

AI music company, named as the kind of creative AI Harrison wants to see more of. "creative things like Suno are are tremendously interesting." 00:58:10

Jev (as transcribed)

Company Harrison says "basically invented a new model category, are associated with it, and have a talented team that can execute on that." 00:52:44 Also named as a brand that rode a viral wave. 00:51:03

AutoGPT

Cautionary tale: fastest-growing GitHub project ever, little heard since. 00:52:04

SaaStr

Referenced via Jason Lemkin on enterprise concerns. 00:43:56

4. People Identified

Miles Grimshaw

Then at Benchmark; led the investment in LangChain and introduced Harry to Harrison. Harry: "I remember Miles telling me about you when he was at Benchmark, and just singing your praises." 00:04:12 Harrison: "when Miles and really the whole Benchmark partnership invested in us, it was me and Ankush. We were two people." 00:53:14

Ankush Gola

Harrison's co-founder at LangChain. "It was me and Ankush. We were two people." 00:53:14

Lin Qiao

Founder of Fireworks AI; co-creator of PyTorch. "Lin and Fireworks, right? They wrote PyTorch." 00:49:47

Alex Atallah

OpenRouter founder, had been on 20VC. "Alex is great." 00:18:05

Eric Vishria

Benchmark partner. Harrison credits him with "Everything just works." 00:47:05

Thomas Tunguz

Investor, credited for the observation that the inference market is huge. "Thomas Tunguz as well had something around just like the market for inference is just so big." 00:47:05

Patrick Collison

Stripe cofounder. Harrison cites his view on building in unsexy areas with low AI penetration and high domain expertise. "It probably wouldn't be like a sexy area... where technology maybe or at least AI penetration is low, but there's a lot of domain expertise built up there." 00:49:14

Amjad Masad

Replit founder; the podcast where Collison made that point. 00:49:14

Sonya Huang

Sequoia partner. "Sonya from Sequoia, who is one of the most plugged-in people in this space." 00:53:53

Tom (IVP)

Led LangChain's Series B after a 2.5-year relationship. "I had been chatting with Tom from like I think right around the Sequoia round, for like over like 2 and 1/2 years." 00:56:00

Bill Gurley

Benchmark; Harrison's pick for coolest to hang with. "Bill Gurley." 00:56:37

Mike Cannon-Brookes

Atlassian CEO, who called many "own-model" claims "largely BS." 00:13:13

Jason Lemkin

SaaStr founder, cited on enterprise attitudes to Chinese models. 00:43:56

Mikey Shulman

Suno's CEO (Harry: "I remember having Mikey on the show"). 00:58:26

5. Operating Insights

Three-step playbook to control coding-agent spend

Visibility, then controls, then harness-level efficiency. "First is basically just visibility. We tracked all of our coding agent usage to LangSmith so we could tell who is spending how much... Second is just basic cost controls. We set kind of like a budget, and if you hit the budget, you can talk to our VP of Engineering and he'll almost always like increase it. But it's just a safeguard... And then third... we control the harness." 00:38:28 Harrison calls it "a pretty repeatable three-step process." 00:39:16

Spend anomalies are usually setup mistakes, not misuse

"We had one or two tasks where some plugin... would screenshot or take a snapshot of the DOM of a webpage and send that back every turn of the agent... I think we spent 40K in a week on that one thing." 00:40:02 The lesson is that observability catches expensive loops that trust-based cultures miss.

Cost-control at the API level: dynamic tool changes without busting the cache

"They introduce things where you can change the tools that are available to an agent mid-flight. Previously, this would have busted the cache and caused like a massive increase in cost. Now, you can just add those dynamically and keep the costs the same." 00:37:08 Prompt caching and harness-aware routing are now core cost levers.

Build the outer loop: run an agent over your traces

For each target (memory, harness code, or model weights), use traces to generate improvements, then gate them with evals. "For bringing things back to the context, or memory, you're probably going to want to run like another agent over those traces... for bringing... to the harness, you're probably going to want to run an agent over those traces and update the code of the harness itself." 00:23:57 Evals confirm whether a change is "actually better." 00:24:26

Evals: trust internal domain evals, distrust academic benchmarks, and read the traces

"I think all the models that come out are kind of benchmark-maxed against those evals, and so you don't learn a lot." 00:21:45 And: "one of the most important things to do is just look at the data... Don't rely too much on kind of like a single number." 00:22:16 Evolve your benchmarks as models improve in different ways.

Fundraising: build relationships years before you need the round

"I had been chatting with Tom from like I think right around the Sequoia round, for like over like 2 and 1/2 years... when it came time to fundraising, I feel like there was like less games in involved there." 00:56:00 He also credits Benchmark with a collaborative, low-posturing approach, and notes VC firms with talent networks deliver real recruiting value. 00:54:30

Three prerequisites to look for: talent, speed, brand

"One, ... the talent that they have is tremendously strong... Two, you have to go like really really fast... And then the last thing would be brand." 00:49:47 On brand: "I think we had a brand advantage that was built up, and that I think that was a durable advantage." 00:51:03

6. Overlooked Insights

Org design may decide how big coding-agent spend gets, and the labor effect is ambiguous

In a throwaway answer, Harrison describes a possible equilibrium where agent spend rises but total outlay stays flat: "the engineers who can do that are incredibly valuable. And so, maybe it actually will just stay the same because the people who can actually do that and who can drive all those agents, there's fewer people, but you're paying them a lot more, and so the spend actually stays the same." 00:41:56 The implication is that token spend may substitute for headcount without reducing total engineering compensation, with value concentrating in a scarce class of "agent drivers." That is a more nuanced scenario than either "AI replaces engineers" or "spend explodes", and it matters directly for how any model-provider valuation is modeled.

The "agent-to-agent" model quietly solves the ad-revenue and load problems that agent traffic creates

Harry raised two objections to agents using sites: lost advertising and reduced basket size. Harrison's reply, delivered quickly, is that a provider-owned vertical agent can embed ads and upsells in natural-language responses and is "going to be better and more efficient at interacting with the underlying systems than a third-party agent will." 00:35:13 This reframes the web's monetization question: ad inventory and cross-sell migrate from pixels to agent responses, and each incumbent app has an incentive to ship its own agent, creating a new product category (vertical agents as APIs for horizontal agents) and likely a large market for the tooling behind it.