Harrison Chase (Co-Founder & CEO of LangChain) has rebuilt his product stack four times since 2022, chasing models that kept changing underneath him. In this episode of Navigators, Paul sits down with him to talk about what LangChain got right, and the one piece of infrastructure it built years too late.

Harrison admits the team underestimated how massive the demand for raw inference would become, and how big general-purpose coding agents would get. He walks through the shift from LangChain to LangGraph to DeepAgents, why sub-agents rarely work the way people expect, and why LangChain now runs its own hiring, content, and coding work through internal agents exposed in Slack.

Topics covered:

• The gateway LangChain should have built a year earlier

• Why agent capability converging on human capability changes the tools debate

• Code mode: why agents writing code beats agents using human-style tools

• Running GTM, content, and coding agents entirely inside Slack

• Why sub-agents need to own whole tasks, not small pieces

• The agent identity problem: delegated access vs. service accounts

• Why LangChain's own AI slop problem got bad enough to need a Slack emoji

  
  

Navigators is a series by Browserbase where we sit down with AI leaders navigating the future of what's happening in AI today.

Host: Paul Klein (@pk_iv), founder, Browserbase

Guest: Harrison Chase (@hwchase17), co-founder and CEO, LangChain

Browserbase: https://browserbase.com 

LangChain: https://langchain.com

Immad Akhund (Co-Founder & CEO of Mercury) uses what he calls the cosmic cube to decide what Mercury builds next: one axis for customer type, one for product, one for compliance and infrastructure. In this episode of Navigators, Paul sits down with him to talk about why that space is functionally infinite, and why trying to expand into all of it at once is how companies lose.

Immad also explains the two-year mistake of waiting to launch a corporate credit card, why he treats every new product like a fresh seed-stage startup inside Mercury, and why talent, not capital, is the real limit on how many bets the company can run at once.

CHAPTERS:

00:00 Why 1-on-1s stopped working once he was managing execs

07:34 Agentic cards: letting AI spend on your behalf

11:09 Why fintech is three times harder than software

13:50 Chewing glass: getting bank-charter ready

16:08 Reaching the edge of knowledge in six months

28:26 Killing 1-on-1s and what Jensen Huang has to do with it

36:39 The "cosmic cube": how Mercury picks its next bet

---

Navigators is a series by Browserbase where we sit down with AI leaders navigating the future of what's happening in AI today.

Host: Paul Klein (@pk_iv), founder, Browserbase

Guest: Immad Akhund (@immad), co-founder and CEO, Mercury

Browserbase: https://browserbase.com 

Mercury: https://mercury.com 

Shawn Wang, aka Swyx, has spent the last several years narrating the AI boom through his podcast Latent Space and the AI Engineer conference. In this episode of Navigators, Paul sits down with him to talk about how far agents have actually come, and why most of the economy hasn't caught up yet.

Swyx points to OpenAI's own research showing that roughly 80% of US GDP could be automated by agents today. The holdup isn't the models, he argues, it's everything downstream: cost, reliability, and a hardware shortage severe enough that people are scavenging memory out of old MacBooks to keep up with demand. He also unpacks the split between "agent labs" chasing domain specificity and "model labs" chasing AGI, and where he thinks the next wave of AI begins.

Topics covered:

• Why 80% of GDP could be automated by agents right now

• The real reason adoption is lagging behind capability

• Agent labs vs model labs: Cursor, Sierra, and domain-specific bets

• Pre-training, mid-training, and post-training, explained simply

• The chip and memory shortage nobody's talking about

• Why "diffusion is cope" divided the AI industry

• The self-driving car tradeoff coming for agents

  
  

Follow Navigators:

https://x.com/pk_iv 

https://x.com/browserbase 

Follow Swyx:

https://x.com/swyx 

Latent Space podcast:

https://www.latent.space 

This data turned AI marketing founder, Kashish Gupta, thinks the reason AI hasn't taken over marketing the way it's taken over coding comes down to one word: taste. In this episode of Navigators, Paul talks with him about what it actually takes to teach a model a brand's aesthetic, and why most companies trying to do this are being lazy about it.

He explains why marketing lacks the verifiable feedback loop that made AI adoption easy in customer support, and how his team spent over a year building the context layer needed to translate a creative director's unspoken instincts into something a model can act on.

He also talks about the subagent whose only job is finding ideas a brand's competitors haven't tried yet, and why he believes the next leap in AI marketing depends less on smarter agents and better prompts, and more on models that generate images and video without inventing details that were never there.

Topics covered:

• Why marketing lags coding: no verifiable loop, and taste is subjective

• Turning a creative director's unspoken instincts into something a model can query

• The subagent built to find whitespace competitors haven't touched

• Why enterprise customers, not SMBs, were the only ones who could adopt this

• Going from access to insights to content: how one problem led to the next

• Where the next inflection point comes from: instruction-following and image generation

• How coding agents changed the day-to-day of running the company itself

  
  

Navigators is a series from Browserbase where we speak with AI leaders navigating the future of building with AI.

Guest: Kashish Gupta, co-founder and co-CEO of Hightouch

Host: Paul Klein IV, founder and CEO of Browserbase

Hightouch: https://hightouch.com 

Browserbase: https://browserbase.com 

Michael Grinich (CEO & Co-Founder of WorkOS) has spent seven years making software enterprise ready, and the last two watching AI companies become his biggest customers. In this episode of Navigators, Paul sits down with him to talk about what's still broken at the front door of every AI product.

Michael makes the case that coding agents already choose your vendors for you, they just can't finish the job: signup forms are built to block automated traffic, so a human still has to step in and register the account. He also gets into why MCP is still growing despite the hype cycle moving on, why he thinks SDKs are becoming unnecessary as agents write their own, and why the next wave of growth will come from agents, not people, filling out your forms.

Topics covered:

• "Stripe for enterprise features": what WorkOS actually sells

• How an SSO and SAML company ended up as AI infrastructure

• The biggest mistake founders make: staying in the pre-PMF experimentation mindset

• Why nothing works unless the management team is AI pilled first

• auth.md, the missing front door for agents

• Is MCP dead? The higher-intent signal hiding in MCP connections

• Why SDKs are going away and coding agents write their own

Yesterday, Airtable got acquired for $1.285B by Bending Spoons. What people don't know is that Hyperagent spun out into it's own company. Howie Liu is doubling down on Agent-first software, and helped build the MVP himself.

On this episode of Navigators, filmed just last month, we sat down to talk about how the Airtable team is rebuilding itself around agents through Hyperagent, why the product form factor has to structurally change with every leap in model capability, and the lean, agent-first companies being born out of all of it.

⁠00:00⁠ Why every model leap forces the product to change shape

⁠02:51⁠ Spinning Hyperagent out of Airtable

⁠05:56⁠ Why Hyperagent bets on horizontal, against the conventional wisdom

⁠10:45⁠ OpenClaw and the overnight-agent stories

⁠13:34⁠ Hyperagent is to OpenClaw what the Macintosh was to Linux

⁠14:57⁠ Why Howie built the Hyperagent MVP himself over the holidays

⁠18:18⁠ "Harness Squared": why the real value lives above the model

⁠28:57⁠ Hyperagent and the rise of lean, agent-first companies

Checkout Navigators, our new pod on Youtube, Spotify, and Apple Podcasts!

Most people still use AI as a glorified search engine. Edward Kim (Co-Founder and Head of Technology at Gusto) thinks that's the blank canvas problem, and it's what's slowing AI adoption across America.

In this episode of Navigators, Paul Klein sits down with Edward to talk about the blank canvas problem, what AI adoption really looks like in the real economy, and what it takes to build a SaaS company today.

They also talk about Gusto Cofounder, a new AI agent that runs the operational work of a small business and only checks in when it needs you.

CHAPTERS:

00:00 Intro

03:39 Prototyping a product in an airport with Claude Code

06:06 Shipping in 10 weeks with no Jira and 9-minute code reviews

10:05 The blank canvas problem and why it stalls AI adoption

12:09 The headless future: running your business by text

13:50 Why 99.9% of people still use AI as a search engine

15:21 Building AI for the small businesses ChatGPT forgot

25:22 Why everyone on the team is a builder now

Browserbase gives AI agents a browser to operate in. Navigators is our series of conversations with the people building the agentic future.

Checkout Navigators on Spotify, YouTube and Apple Podcasts.

Aaron Levie (Co-Founder & CEO of Box) has run Box through three platform shifts: cloud, mobile, and now AI. In this episode of Navigator, Paul sits down with him in Redwood City to talk about what's actually changing as agents take over knowledge work.

Aaron makes the case that most "AI problems" are actually context problems. Agents need access to the 90%+ of enterprise data that's unstructured, and giving them the right information in a tiny window of time and space is the real unsolved challenge. He also breaks down why there's a growing gap between model capability and real-world adoption, why the applied layer will stay valuable even as frontier models get more powerful, and why founders building today need to treat go-to-market as the new scarce resource now that building software is cheap.

Topics covered:

• Why agents use data more like people do than computers ever did

• The "model overhang": why capability is outpacing adoption

• Why Silicon Valley will always move faster than the real economy, and why that's fine

• How Box is rebuilding its own ops around agents internally

• Headless SaaS, agent-native pricing, and what happens when agents outnumber humans 100 to 1

• Why the applied layer wins even in a bitter lesson world

• Aaron's advice for founders starting companies from scratch today