Agent Adoption

Hypotheses behind my journey leaving Google and starting an agent startup

I left Google to start a startup because I believe Agents are capable enough to drive large-scale industry transformation today, but mainstream consumer adoption is blocked by the application layer: too much setup, too much reliance on chat, too little daily utility, and not enough human centricity. The products we have today aren’t ready.

The companies that make agents invisible, reliable, and useful every day will define the next era of software. These are the hypotheses behind Vycari.

1. We are very early in a complete transformation of the tech industry

I learned HTML in 1998 and gleefully applied the <blink> tag liberally. We’re at the same point in the adoption cycle with agents today as we were with web then: a lot of hype, some early public victories, uneven distribution, but ultimately most people aren’t using agents. Most work hasn’t been impacted. There are obvious exceptions, but outside of well-publicized enterprise or software engineering focused projects, agents are largely confined to hobbyist projects.

We have a lot of Homebrew Apple-Is and no iPhone - hobbyists are full of hype but there’s no breakout consumer success.

It may not be immediately visible what the gaps in the industry are with the future transformed state, but they exist everywhere - the simplest example I can provide you is I can’t add an agent to a video call and ask it live to do work on my behalf. It’s been clear that’s a mainstream capability we want for over a year and it’s not broadly available.

Big company velocity is largely untouched as well - code is generated faster, but reviews are slow, alignment is slow, shipping is slow.

We’re still in early days. Agents will transform everything.

2. Today’s agents are hobbies, not products

As Allen wrote, Not Everyone Wants a Hobby. Agents today require too much work and maintenance to be adopted broadly.

The future is not here yet. Everything is brittle and there are edge cases everywhere. Configuring an agent takes smart people a lot of time. For a consumer experience, a lot of time is more than two clicks and five seconds. Anything longer is too much.

Among my other jobs, I founded the Google Photos data team and we spent a lot of time trying to suss out what drove retention. The only thing I can tell you I KNOW that increases retention is decreasing cold start latency. Everything else was mixed.

Cold start latency for using an agent is measured in minutes to hours with many unnecessary profanities today. And even when you get it working, it’ll break or do things you don’t expect. You can’t build scaled consumer software on stuff that breaks this easily.

Sure, if you know what an MCP is and what the magic words are you can do anything.

Most people don’t want to learn magic, they just want to get something done. Most people are not using agents. Most people can see we’re failing them at large today.

This includes our obsession with vibe coding as a panacea for everyone. Vibe coding isn’t for everyone. I love vibe coding and I have been doing it for a long time, except at some point I stopped prompting humans and switched to agents.

I appreciate it is apocryphal, but as Ford was reported to have said, most people just want a faster horse. They don’t want to spend their lives agonizing about layout choices or how to improve a workflow. Most people don’t know what is possible with the current technology, and don’t know what to ask. They have lives to lead. We do not need burden them with unnecessary cognitive load.

We need to get out of the way and turn agents into durable products, not hobbies.

3. Chatbots are a path to the future, not the future

Chat is not the end state of an agentic transformation of software. It’s winning today because of the flexibility it offers, but most people don’t want to do most things conversationally. Most workflows are much better expressed with multi-modal capabilities.

I led a redesign of the Google Assistant in the dinosaur era of 2017 and the clear learning we processed back in the stone age was that text is not sufficient for users. Consider any shopping journey for example. Consider timers, alarms, date pickers - all are best represented by elements that are superior to typing in text. For example, “No, change the alarm to 6:15” vs. scrolling a selector.

Before you say “but voice is easier than your petty UIs!” - I am a voice maximalist, and believe in its future, but try guiding your friend on how to build a lego using only verbal instructions without pointing or gesturing. We did a version of this as a team activity once and it’s a lot harder than you think! Voice as an isolated interface isn’t as rich as a cohesive multimodal user experience system.

One problem people have today imagining the future is it’s not clear to us what the UI of the future looks like, but chat works okay and so we put chat everywhere. Chat is great, I love chat, but a dynamic UI is essential.

4. Agent interfaces must adapt to the task

Not all UIs should be chat - we need more flexible UI systems that adapt to divergent use cases. We do everyone a disservice by narrowing all experiences into a text stream.

Consider the deficiencies of a chatbot when determining what to do when you open the app - it is completely dependent on the user having a known prior intent. The blinking cursor is a real problem. There are many paths to helping people know what magic spell they can utter to get a response, but the ultimate best path is a rich UI - just show people what there is vs. making them think about what to say. We have exceptional ability to process rich visual information, and the path to success is using agents to lean into that capability, not sitting back in the swamps of text chat. Just look at all of the complaints about the verbosity of chatbot outputs!

Open a physical copy of the New York Times in 1996 and compare it to their website at the time and it looks very similar - we just slapped a newspaper on a website. Newspapers on websites have evolved since then - at some base level it’s primarily columns of text so the fundamental nature of the offering hasn’t changed, but everything about how it is generated, distributed, and offered has changed. It has also evolved incredibly rich capabilities that transcend text by interleaving visuals & audio with a textual format. Someone from 1996 would be just fine adapting to our modern newspapers, but boy they would also be happier because it is legitimately better with adaptable rich UIs.

It took years for the modern online news formats to emerge. We need to be actively pushing the boundary on what UIs look like in the world of agents. Much like our 1996 time traveler, if you time jump ahead 10 years from now you’ll adapt just fine into the new world that is coming - fundamentally we will still be looking at screens (with greatly enhanced wearables) but the formats will be far superior to what we have today.

People still want to read, to look at diagrams and pictures, to watch video, so the fundamental nature of content will be the same. We’ve lost a bit of that grounding in our thinking about design systems. There are some historical constraints in web design today that apply due to the ads business model forcing us into suboptimal design choices, but those constraints are being lifted as the ecosystem is forced to adapt to a world where singular interfaces threaten to eat the web.

Ultimately we are humans and we have great visual acuity and our interfaces will continue to take advantage of that with increasingly rich and personalized visualization.

5. Adoption begins with existing daily needs

Before I went to Google I worked at Bump, an early mobile darling. We had an incredibly viral app - you used it to easily exchange phone numbers with others, removing the friction around typing in a number. Viral by nature, it was a huge hit.

Unfortunately it was not retentive by nature because the average person has to exchange a phone number something like once every three months. Using an app every three months doesn’t get you sustainable growth and there’s no real business or product value you’re driving, but if you’re lucky you get acquired by Google and build Google Photos. Just kidding, not just lucky, lucky and good.

But I want to impress on you the essentiality of a daily driver use case. You need to build experiences that pass the toothbrush test as Larry Page used to put it. We need to build experiences that form a baseline of daily engagement, and from there you can build a large variety of use cases that people use once every three months, but only if you have engaged people in their daily lives.

Toothbrushes are not flashy! You aren’t going to win 30 under 30 awards building toothbrushes! But toothbrushes are essential. The daily experiences we build might look superficially very similar to what exists in a pre-agent world, but the core logic will get a massive upgrade in both power and usability thanks to agents.

A focus on daily use cases also requires a focus on trust: you must create something reliable enough that it engenders trust and reliance. We cannot get users to adopt based on vibes alone, we need to build trustworthy experiences. The prime directive in building trust is centering humans in our experiences. People are the top priority.

I believe it remains possible to build experiences that are fundamentally transformative, but begin with daily essentials which extend into an increasingly large variety of use cases.

6. The application layer is wide open

The model is not the product, it is one component in a powerful stack of primitives. Value is moving up to the application layer. There are greenfield opportunities for growth everywhere.

We are in an age of transformation. Much like the Cloud boom kicked off by AWS, a lot of the transformation has been invisible to the world at large - as things moved to the Cloud users didn’t notice much other than things got cheaper and more reliable. Unfortunately we’ve gotten off to the wrong foot with this transformation with rampant fear mongering and self-serving predictions of doom.

There’s enough success for all of us. We can create a world of increasing abundance including within the tech industry itself. The application layer is wide open. It’s time to build. I am excited to share what we have with you soon.

If you’re interested in hearing more, sign up to be an early tester here.