02 / Made with
TC 00:01:45:00
JK · Reel 2026
The intro you just watched
The voices
Gemini 3.8 Flash TTS
Shipped last week · a host and a skeptic, directed
The reel
Claude Opus 5.5
One prompt
Submitted · Feb 2026
The question hasn't changed. My answer has.
Enjoyed.
voices = Gemini 3.8 Flash TTS, launched last week. Provided script, picked voices, gave acting direction.
animation = 2 prompts to Opus 5.5 (the conversation, then resume reel). stitched together - ok, my agent did
CLICK AI commentary right - submit in Feb
started pulling the talk together - struggled - description didn't cover what's top of mind
question hasn't changed. answer has
"year of agents" for a while... recent moment really hit home
CLICK
I hope you enjoyed watching that half as much as I enjoyed making it.
The voices are Gemini 3.8 Flash TTS, our new text-to-speech model, which launched last week. I gave it the script, chose the voices, and gave them their acting direction. The animation took two prompts to Claude Opus 5.5: one for the conversation, one for the resume reel. Then I stitched them together (I mean my agent did).
CLICK And that AI commentary track is right - I did submit this talk in February. So when I recently started trying to pull the full talk together, I kept struggling, because the description I wrote for the talk back then, didn't really touch on the most topical things I felt worth discussing today.
Luckily, the main question hasn't changed. It's still an important one to be asking...but my answer has.
I feel like we've been hearing that this is the year of agents for a while now. But it really hit home in a recent moment for me.
A podcast I made with an agent
Maya and the Fog
Scary stories for brave kids. Maya is closely modeled on my daughter, Ophelia.
Episode 1 ~24 hours · coached segment by segment
started a side project w/ Ophelia (5): scary story podcast. AI writes it, ElevenLabs narrates
CLICK play clip, ~11s
episode 1 = ~24 hrs. built in segments: build, listen, give notes, skill (dialogue/pacing/voice/sound), iterate, next segment
CLICK episode 2 - same start. segment 1... then Ophelia looks up CLICK
You see, I started a side project with my five-year-old daughter, Ophelia. It's a scary-story podcast called Maya and the Fog. AI helps write it, and ElevenLabs narrates it.
CLICK The clip plays, about 11 seconds. Let it finish.
Episode one took us about 24 hours. I built it in segments. I'd have the agent build one segment, then we'd listen to it, and we'd give it notes. Dialogue, pacing, voices, sound. It would take the feedback and iterate, and when we were finally happy with a segment, we'd move on to the next.
CLICK Episode two started the same way. The agent built segment one, and we sat down to listen. And when it finished, Ophelia looked up at me and said:
“Can you tell the agent to make the rest? ”
Ophelia, age 5
AI isn't something you talk to.It's something that does things.
no notes. neither did I. agent made the rest
CLICK important for 2 reasons
1: built + launched full podcast on Spotify under 24 hrs - impossible year ago…6 months
2: to daughter, AI isn't something you talk to. it's something that DOES things - builds, makes, creates, acts…EVERY USER
Let the line on the slide land. Beat.
She had no notes - and neither did I. So I had the agent make the rest.
CLICK And this moment felt important for 2 reasons:
1. I was amazed I could build a full podcast and launch it in Spotify in less than 24 hours (that would have felt impossible a year ago....even 6 months ago)
2. But perhaps more importantly - because it made me really realize that to my daughter, AI isn't something you simple talk to. It's something that does things for you. It builds, it makes, it creates, it takes actions.
That's the expectation she's growing up with. It's my expectation now too. And it will be the expectation every one of your users in the (not too distant) future.
When making things gets this cheap
Can you build it?
Should you build it?
Two parts
01 What's worth building
How I decide.
Three forces Four pieces of advice
02 How to build it
Once you've decided.
Two users Bigger than your UI Someone else's agent
models have gotten incredibly capable. when making is this cheap,
CLICK question isn't "can you build it" - it's what's worth building
CLICK 2 parts: decide what's worth building, then how to build it
The models have gotten incredibly capable. And when making things gets this cheap,
CLICK the question isn't "can you build it?" anymore. The question is: what's actually worth building?
CLICK So I'm going to split this talk into two parts. First, how to decide what's worth building. Then, once you've decided, how to think about building it.
The three forces
Together they decide what's possible, and where things are going.
01 Models + the harness
Jumps weekly, sometimes daily
02 The ecosystem
And so is the product value prop
03 Users
Down to UX patterns and product shape
They don't take turns. They all move at once.
3 forces - together decide what’s possible - where going
CLICK 1: models + harness. dramatically better - (tools, memory, the ability to act). possible jumps weekly. assume last month's ceiling you're behind
CLICK 2: the ecosystem. isn’t people-to-chatbot anymore. agents using products, products deciding. permissions, integrations, biz models - the contract of internet being renegotiated. value prop shifts too
CLICK 3: users. Trust agent keeps shifting. Overton window moves. some lean in, some push back - slop. baseline normal moves - my expectations today aren’t the ones I had a year ago
CLICK can't pick just one. all moving at once - I'm constantly calibrating. advice in 4 parts
I start with three forces, because together they decide what's possible, and where things are going.
CLICK First: the models and the harness. The models got dramatically better. And so did the harness around them: the tools, the memory, the ability to act. The thing that gives a model hands. What's possible jumps weekly, sometimes daily. If your product assumes last month's ceiling, you're already behind.
CLICK Second: the ecosystem. It's not just people talking to chatbots anymore. Agents are using products, and products are deciding whether any agent can use them, or only their own. Permissions, integrations, bots, business models. The contract of the internet is being renegotiated in real time. And so is the value prop of a product.
CLICK Third: users. What people will trust an agent to do keeps shifting. The Overton Window is moving. Some people lean in hard. Others push back on slop and fakes. And the baseline of normal moves with them, right down to the UX patterns people expect. My daughter's expectations are not the ones I had a year ago.
CLICK The hard part is...you don't get to focus on just one force while ignore the others. These are all moving at once - and I find myself constantly callibrating. So with those in mind, here's my advice for deciding what's worth building, in four parts.
Advice 1 2 3 4
Advice 01
Push the frontier.
Use the frontier models, and push them to their limits.
One prompt ← the showreel
The other end GPT-6 Astra, driving Blender · many, many rounds
1: use frontier models. push them to their limits
Intro - example: showreel = one prompt. here's another
CLICK dancer plays - GPT-6 Astra driving Blender. lots of iterations, still not "done" - but it’s impressive
One: use the frontier models, and push them to their limits.
The showreel you just watched is one example. That was one prompt.
Here's another example.
CLICK The dancer plays. This is OpenAI's GPT-6 Astra, driving Blender. This one took a lot of iterations - and still isn't done - but I'm really impressed with what it's gotten to.
Advice 1 2 3 4
In the last ten minutes, you've seen
Three labs. Three models. Three weeks.
Sep 3
GPT-6 Astra
OpenAI · the dancer
Sep 22
Claude Opus 5.5
Anthropic · the reel
Sep 23
Gemini 3.8 Flash TTS
Google · the voices
Today · Sep 30
Last 10 mins, 3 examples, 3 model launches, 3 weeks. Already outdated! That’s the pace possible.
knowing what's possible shapes what you build
So in the last ten minutes you've seen work from three frontier labs, prototypes and presentations built on three recent model launches, all from the last three weeks.
That's the pace of what's possible right now. And knowing what's possible is what shapes what you should build.
Advice 1 2 3 4
My side projects
Things I build because I want to.
Two coloring books
On Amazon
My salad cookbook
Finally printed
A mermaid game
For the iPad
Interactive website
mayaandthefog.com
A scary-story podcast
Maya and the Fog
A dancer
GPT-6 Astra + Blender
This intro
Gemini TTS + Opus 5.5
None of them needs to exist. They're an excuse
to learn what's possible right now, to find where it breaks, to make connections, and to build product taste and conviction.
Keep side projects: 2 coloring books, salad cookbook, iPad mermaid game, Maya + interactive site, a dancer, my intro... to name a few
CLICK none are big. things I like. Excuse to build, learn possible, what breaks. Make connections (not) reading.
that's how I build taste & conviction. build to learn.
That's why I keep so many side projects going. Two coloring books. My salad cookbook. A mermaid game for the iPad. The Maya and the Fog podcast, and its interactive website. A dancer. And the intro you just watched...to name a few
CLICK None of these are big things. They're things I like. But really, they're an excuse. An excuse to learn what's possible right now, to find where these systems break, and to make connections I'd never make just reading about them.
That's how I build product taste, and conviction. I build to learn.
Advice 1 2 3 4
Then project it out
Build for where the models are going.
LENGTH OF TASK A FRONTIER MODEL CAN COMPLETE ↑ 0 12 HRS 1 DAY 2 DAYS 3 DAYS 2023 2024 2025 2026 2027 GPT-4 o3 GPT-5 Claude Opus 4.6 Claude Mythos Preview TODAY’S CEILING Today: you start building You launch Where the models will be Data: METR time horizons, 50% task success, May 2026. Linear scale. Line: fit to 2025-26 frontier models, doubling about every 3.7 months.
Key: don't stop at possible today. project out
Chart tracks how long a model can complete a task - 2023. Progress speeds up.
best way to feel where going: stay curious, build for fun / intuition
CLICK build for today's ceiling - launch outdated. build for where models going
goal: ride the tailwind - better models should make product better (even if rewriting)
But the key is: don't stop at what's possible today. Project it out.
Point to the chart. This chart tracks how long a task a frontier model can complete on its own, starting with GPT-4 in 2023. The rate of progress continues to speed up.
CLICK So if you're building at today's ceiling, by the time you launch, you're already outdated. Build for where the models are going. And the best way to internalize where you think they are going, is to stay curious and build for fun, develop that intuition.
The goal should be to ride the tailwind. Every time the models get better, your product should get better too. Even if that sometimes means rewriting what you've built.
Advice 1 2 3 4
Advice 02
Strip the guardrails.
The scaffolding you needed to make it work is limiting what the model could do now.
Model
Pipelines
Strict constraints
Special handling
Predefined steps
Hand-picked models
Retries
Validators
Fallbacks
Fixed formats
Dead weight
2: over time, strip the guardrails
built guardrails early to make work: pipelines, strict constraints, special handling. scaffolding. used to spend quarters getting one model good at one use case
CLICK as models improve, scaffolding dead weight. not protecting you - limiting what model could do
a pattern that's come up again & again last few years
Which brings me to two: over time, strip the guardrails.
Even if you've built a great AI product, there's a good chance that when you first built it, you had to put a lot of guardrails on it just to make it work. Pipelines, strict constraints, special handling. A whole lot of scaffolding to force the model to behave. We used to spend quarters making a model good at one vertical that way.
CLICK As the models get better, a lot of that scaffolding becomes dead weight. It's not protecting you anymore. In fact, it's limiting what the model could do, if you let it.
This is a pattern that has come up again and again as I've been building products over the last few years.
Advice 1 2 3 4
A workflow builder.
It started by chaining steps. For each one you picked the model, set the instructions, selected the tools.
Pick the model Set instructions Select tools
A deterministic workflow
Opal · Google Labs · build, edit and share AI mini-apps
ex. Opal: a workflow builder
chained steps - pick the model, set instructions, select the tools
Launched as deterministic workflow that leveraged AI
Here's an example: Opal. Opal is a workflow builder. It started with you chaining steps together. For each step, you picked the model, set the instructions, and selected the tools. You created a deterministic workflow that leveraged AI.
Advice 1 2 3 4
February 2026: the agent node
At launch: every step wired by hand
1 Ask the user three questions Predefined
2 Research the topic Web search ▾
3 Write the report Gemini ▾
4 Make a video Veo ▾
5 Lay it out 4 pages
That structure was the guardrail. It made the output reliable.
The agent node
You give it the goal A report on this week's top social trends, with a short video.
Web search Gemini Veo Asks you what's missing
It was right for the models we had. It isn't right for the models we have now.
How much of what you built can you delete and hand to the model?
Feb: launched "agent node." collapsed - no longer wire node - one goal, agent does the rest
CLICK not a knock on v1 - right for models at launch. models evolved, we had to too
CLICK advice: ask "how much can I delete and hand to the model?" loosen guardrails as models earn it
Then in February, we launched the agent node. We collapsed all of that. Instead of wiring up multiple steps, you have one node with a goal, and the agent figures out the rest.
CLICK That's not a knock on how we built it the first time. It was right for the models we had at launch. But the models got better, and things evolved, so we had to evolve too.
CLICK So my advice: every so often, look at what you've built and ask, "How much of this can I delete and hand to the model?" Loosen the guardrails as the models earn it.
Advice 1 2 3 4
Advice 03
Keep the lessons, not the prototype.
Sometimes the whole approach, or the whole product, stops making sense.
Keep
The lessons. They carry into everything next.
Turned off · Mon, Sep 28
Mixboard · Google Labs · an AI concepting board · ran about a year
3: sometimes guardrails isn’t enough - whole approach, or product - rethought.
2 days ago: shut down Mixboard. Google Labs experiment - concepting board, ran a year.
CLICK shutting down hard - sunk cost. but keeping it is bigger cost - opportunity cost, build something better - only now obvious.
UPSIDE: Learnings carry forward. (user behavior, UX, model capability). Apply that to next.
Advice: keep the lessons, not the prototype.
Pivoted++ !=start then fresh start is best - not one more patch
Three: sometimes rethinking the guardrails isn't enough. Sometimes it's the whole approach, or the whole product.
Two days ago, we shut down Mixboard. It was a Google Labs experiment, a concepting board, and it ran for about a year.
CLICK It's hard to shut things down - sunk cost and all. But keeping it around is more expensive, because it's lost opportunity cost that you could be spending on building something better - something that is only now obvious to you. And the positive side is how much we learned by building it. About user behavior, about UX, about what the models could do. And all of that carries forward.
So keep the lessons, not the prototype. And if you've pivoted so many times that where you ended up looks nothing like where you started, the honest move could be a fresh start, or a different path. Not one more patch.
Advice 1 2 3 4
Advice 04
You still set the goal.
EPISODE 1 · SEGMENT BY SEGMENT EPISODE 2
Your version of good THE GOAL · SET BY YOU First pass: “make the rest”
“This line lands. That one doesn't.”
“This pacing is off for this scene.”
Notes become skills it keeps.
The dancer
“Not that,” round after round.
4: you still set the goal
CLICK everything said = give more to model, loosen the guardrails, delete scaffolding. only works if know what aiming for.
CLICK ex. Maya: built episode 1 in segments - build, listen, iterate w/ the agent, improve skill - new segments start better, less iteration - quality hillclimb - I set goal (my version of good)
by ep 2, 1 seg in, told to generate rest. not perfect - small notes & iterated. but much better starting.
CLICK dancer: same story. v1 ok. iterated round closer to MY goal
Four: you still set the goal.
CLICK Everything I've said so far is about giving more to the model. Loosen the guardrails. Delete the scaffolding. But that only works if you know what you're aiming for.
CLICK Go back to Maya. I built the first episode in segments. Build, listen, iterate with the agent, improve its skills, then build the next segment. And every new segment started out better, with less iteration. It was a hill climb on quality, and I was the one setting the goal. My version of good.
So by episode two, one segment in, I could just tell it to generate the rest. It wasn't perfect. I still had small notes on the full episode, and we still iterated. But it started from a much better place.
CLICK The dancer is the same story. The first version was okay. Then I iterated, round after round, until it got to my goal.
An agent will happily build you something plausible.
An early round Plausible
Many rounds in Closer
Final My version of good
many, many rounds many more The difference between plausible and good is on you.
agent will happily build you plausible
CLICK difference between plausible & good is on you
goes for product AND model output. as models do more of work, right goal and your definition of good are what set you apart.
An agent will happily build you something plausible.
CLICK The difference between plausible and good is on you.
That goes for your product, and for its output. As the models get better, they'll do more and more of the work. Setting the right goal, and defining what good looks like, is what will set you apart.
How to build 1 2 3
The product itself
Every product now has two users.
The person
A real interface: a confident default, and a way to override it.
The agent inside your product
It can do anything the person can, in code.
my mental model has shifted: every product has 2 users - the person, and agent
how I think: anything person can do, an agent should be able to
CLICK just for fun - this deck has 2 users too. press A for agent mode. here's what it sees
My mental model has shifted. Every product now has two users. There's the person. And there's the agent: the AI that's also using your product.
The way I think about it now: anything a person can do in your product, an agent should be able to do too.
CLICK And just for fun, this deck has two users too. If you open it and press A, it switches into agent mode. Here's what an agent sees on this slide.
How to build 1 2 3
How much the user delegates · e.g. a party-planning agent
Do It With Me ↔ Do It For Me
It asks you
What tone should the invite have?
Type your answer
It suggests
Pick a tone:
Playful Warm Formal
It assumes and acts
Chose a tone Wrote the invite Sent it
Do it with me Do it for me
Same product. Different people, different points on the line.
OK start with human user
designing AI-native products, I think along spectrum. Do It With Me (user drives, A helps) - Do It For Me (delegation)
ex. a party-planning agent
CLICK DWM side: user initiates, agent asks everything
CLICK middle: agent starts suggesting options to pick from
CLICK far end: agent picks goal, assumes everything, acts
CLICK challenge: designing whole spectrum into ONE product - think about this a lot. Different users want different delegation, sometimes same person on different days.
none of them are wrong. so: how might we design a UX that flexes to that?
Ok, let's start with the human user.
When designing products that are AI native, I like to think along a spectrum. On one end, Do It With Me: the user drives, and the AI is there to help. On the other end, Do It For Me: the user delegates the whole thing to the agent.
Take a party-planning agent as an example.
CLICK On the DWM side, the user initiates the action and the agent would ask it everything.
CLICK Moving to the right a bit, the interaction may be one where the agent starts suggesting options for the user to simply pick from.
CLICK Then on the far end, the agent might pick the goal, assume everything, and act.
CLICK The challenge - designing that whole spectrum into one product - which I find myself thinking about a lot these days. Because different users want different amounts of delegation. Sometimes even the same person might have a difference preference depending on the day. And none of those users are wrong for wanting it.
So the question becomes: how might we design a UX that flexes to that?
How to build 1 2 3
A level deeper
Your product can be bigger than your UI.
Before
Every setting had a place in the UI for the user to select.
Setting Studio
Props None
Time of day Noon
Human presence Off
Location Select...
Mood Neutral
Occasion Select...
Indoor / outdoor Indoor
Now
Setting Studio Sunny terrace
Props None Espresso cup
Time of day Noon Golden hour
Human presence Off Worn by a model
Location Select... Mediterranean
Mood Neutral Relaxed
Occasion Select... Weekend
Indoor / outdoor Indoor Outdoor
a level deeper: product is capable things never exposed in UI
before: every feature needed place in UI so person could use it (why menus & settings exploded)
CLICK now: user simply asks outcome, agent changes whatever settings needed - some settings not worth exposing at all
Let's go a level deeper. Your product can be capable of things you never expose in the interface.
What do I mean by that? Well, previously, every feature had to have a place in the UI for the user to use it (that's why menu bars and settings have exploded).
CLICK But now, the user can just ask for the outcome they want, and the agent can change any settings necessary to get there - and that could mean that some settings simply aren't worth exposing.
How to build 1 2 3
Pomelli · Photoshoot · a real run
One product photo. A few clicks.
1 Your website
Pomelli builds your Business Profile: colors, fonts, tone, aesthetic.
2 Pick a product
Pulled straight from your site's catalog, or upload one.
3 Templates, suggested
Six picked for this photo in one click. Override anything.
labs.google.com/pomelli
Recorded in Pomelli · opulize.com, a London eyewear brand
ex. Pomelli Photoshoot: creates professional photoshoot from a low-quality phone photo
CLICK walkthrough plays - start with website, Pomelli builds Business Profile. select photo. behind scenes we choose best templates (can override). one click generates the images (can edit)
Here's an example from Pomelli: Photoshoot. It creates a professional photoshoot from a low-quality photo taken on your phone.
CLICK The walkthrough plays. You start with your website, and Pomelli builds your Business Profile. Then you select a photo. Behind the scenes, we figure out the best templates to make that product look good (you can always override the choices). Then, in one click, it generates the professional looking images (you can also keep editing from there if you want to).
How to build 1 2 3
What came back
One click on top. A lot of decisions underneath.
On ice Floral whimsy Flat lay Model try-on In context Seasonal
In context → Your Business DNA →
Makeup brand In a purse On a counter …
Sunglasses On a face Pushed up on a head …
60+ contexts we never show as settings We say “in context.” We handle the rest.
Skills go deeper than your UI Build a great UI for people. Also build for the agent.
built way more capability / what's UI. want interface stay clean & understandable, not everything exposed
CLICK deeper on templates: pick best ones & how to apply it. Take "in context" - uses bDNA to define what “context” means. makeup brand: purse / counter. sunglasses: on face / on the head. agent figures it out
CLICK alternative is exposing 60+ options to users. didn't want!
CLICK skills go even deeper. Causes interesting gap…product be more capable than its interface. natural language better way to configure and unlock those features, enabling agents even further.
Lesson: build great UI for ppl, also build for agents
Let's take a look at how that feature works - because we built a lot more capability than what's in the UI, but we wanted the interface to stay clean and understandable - so it isn't all exposed to the user.
CLICK Let's go deeper on the templates. Not only do we try to select the best templates for the product you selected - but we also figure out how to properly apply it. So say you select "in context." We use your Business DNA to understand what "in context" means for you. If you're a makeup brand, the context might be in a purse, or on a counter. If you sell sunglasses, it might be on someone's face, or pushed up on their head. The agent figures all of that out and generates the final photoshoot images.
CLICK The alternative would have been expose sixty-plus options to our users - which we did not want to do.
CLICK The other thing to consider is if you build skills that work with your product, because these can enable even more functionality that what is exposed in your UI. So there's this interesting gap that emerges: the core product can be more powerful than what's exposed on its interface. And natural language is often the better way to configure all those settings to unlock different features.
The lesson to walk away with is this: build a great UI for people. But also build for the agents - who might be even more capable ;)
How to build 1 2 3
The flip side
Which agent? Increasingly, someone else's.
xAI · Aug 2026
Grok Bot
Persistent agents with their own computer, running around the clock.
Startup · Aug 2026
Instinct
Connects to your email, messages and screen, and acts proactively.
Meta · Sep 8, 2026
Muse
A personal agent across Meta’s apps, with your accounts and payments.
Who's asking for web pages now Cloudflare Radar · share of HTTP requests for HTML
Bots · 57.5 %
Humans · 42.5%
Assume agents will use your product.
up until now, focused on agent YOU build into product. BUT consumer agents blowing up
CLICK last few months: Grok Bot, Instinct, Muse, Dots - all launched. more bot traffic / human traffic. Agents on the internet, and in your product, aren't always yours
CLICK advice: build assuming other agents - use your product
Now, Up until this point, I've been mostly focused on the agent you build into your own product. But consumer agents are blowing up.
CLICK Just in the last few months: Grok Bot. Instinct. Muse have all launched. There's currently more bot traffic on the web than human traffic, and that trend is only going to continue as agents enter the picture. And it's important to remember - the agents on the internet, and using your product, aren't always yours.
CLICK My general advice: you should build assuming other agents are going to use your product.
How to build 1 2 3
My Strava · Dog Lake hike · Sep 19
Product decisions are already being made for companies this way. Invisibly.
personal example, few weeks ago: hiking in Yosemite and GPS glitched - line on the map…fixing headache
CLICK normally, as a PM: watch how often people hit a workaround & how much friction it causes - decide if worth fixing
In this case - my agent did it. didn't care. no friction to push through - no signal for anyone to see.
CLICK question: how should Strava fix this? UI, agent skill?
If tracking user trajectories / friction logs to prioritize…data being complicated. invisibly. agents quietly changing what "user behavior" even means.
Here's a personal example from a few weeks ago. I was on a hike, and my GPS glitched. Take a look at the map - can you see the section where there is a straight line that just...juts out....clearly not something we actually did. And because I care about accuracy (mostly because my 3 year old twins hiked this entire thing on their own and I wanted to capture that properly), I wanted to fix it. But the manual correction in Strava was going to be a whole thing. Downloading the file, cropping out the problematic section, re-uploading it - it's a headache....So instead, I sent my agent to fix it for me.
CLICK
Normally, as a PM, you'd watch how often people hit that workaround, and how much friction it caused, to decide whether it's worth investing in the real fix.
What's interesting is that my agent did it for me. It didn't care. It had no friction to push through, so there was no signal for anyone to see.
CLICK Which means I actually don't know if Strava should make this easier now. My agent muddied the data. So it's not just that product decisions are being made for companies this way. They're being complicated. Invisibly. Agents are quietly changing what "user behavior" even means.
How to build 1 2 3
To do that, I handed it my login
What should an agent be allowed to do? Who decides?
One way to think about it: one-way door or two-way door?
Two-way door
You can walk it back.
Fix a route Edit a draft Change a setting
The agent can act.
One-way door
It's permanent.
Purchase Delete Send
A human confirms first.
To accomplish - hand agent login. real questions: what identity should it have? what should it be allowed to do? who decides?
CLICK lightweight think about permissions: one-way door vs two-way door
2 way: can walk back - ok for agent to act.
1 way: permanent - HIL, confirm first
Now to do this, I handed my agent my login. Which raises real questions. What identity should the agent have? what should it be allowed to do? And who decides?
CLICK One lightweight way to think about agent permissions is: it something a one-way door or a two-way door. When an agent acts, Can you walk it back? If so, then it's probably okay for the agent to act. On the other end, is a permanent action. In this case, you probably want a human in the loop, to confirm before it acts.
How to build 1 2 3
Remember the dancer? Made in Blender.
I have never learned Blender.
I'm a Blender user now. Someone else's agent brought Blender a user it would never have reached.
Comes to other agents using products - why it’s exciting
remember the dancer…Blender? I’ve never used Blender, never intended to - it’s powerful, but really complicated. put simply: not "a Blender user"
CLICK except that I am. because my agent is
CLICK for Blender team, a consumer agent gave them a new user
building your product so other agents can use it could grow your user base
Now, here's why agents using products is exciting.
Remember the dancer? The one made in Blender. The most interesting part of that might be that I had never used Blender before, and I had no intention to - Blender is powerful...but it's also complicated....like, really complicated. Put plainly...I am not "a Blender user".
CLICK Except that I am. Because my agent is.
CLICK So for the Blender team, a consumer agent just gave them a new user they would never have reached.
So building your product in such a way that it allows other agents to use it, could be a great way to grow your user base.
How to build 1 2 3
If somebody's agent is going to use your product
Why should your product still exist?
Someone's agent
Another product
One call away
Your product
Why would it choose you?
It builds its own
Anything can be built now
No one-size-fits-all answer Answer it on purpose, before an outside agent answers it for you.
raises question: if someone's agent uses your product, how does that change your value prop? as agents get more capable, why should your product still exist? why would an agent choose it?
CLICK no one-size-fits-all answer. but take the time to figure it out - or it'll be answered for you
But it does raise a really important question - someone's agent is going to use your product: how should that change your value prop? As agents become increasingly more capable, why should your product still exist? And why would an agent choose it?
CLICK There's no one-size-fits-all answer. But you should take the time to figure out what your answer is. Because if you don't figure out the answer, and design that into your product, it will be answered for you.
Worth asking about your product
Three questions.
01 Would you build this today, knowing where the models will be tomorrow?
What's worth building
02 Can your agent do everything your user can?
Your agent
03 When somebody sends their agent, why does your product still exist?
Someone else's agent
3 questions to take with you, thinking about what you're building
CLICK knowing everything you know now, would you build this today, keeping in mind where the models will be tomorrow?
CLICK can an agent do everything your users can?
CLICK when someone sends their very capable agent, why would they choose your product? why does it still need to exist?
So, three questions to take with you as you think about the products and features you're working on:
CLICK Knowing everything you know now - would you build this today, keeping in mind where the models will be tomorrow?
CLICK Can an agent do everything your users can?
CLICK And when someone sends their very capable agent, why would they choose to use your product? Why does your product still need to exist?
Pause.
