# AI Valuations, Supply Chain Shifts, and Agent Workflows

**Podcast:** The AI Daily Brief (Formerly The AI Breakdown): Artificial Intelligence News and Analysis
**Published:** 2026-05-11

## Transcript

Today on the AI Daily Brief, what is the best way to be talking to our agents?
Are we using Markdown, HTML?
Believe it or not, that debate has been raging, and I think it's for a bigger reason than it seems that has to do with a big shift in how we work.
Before that, in the headlines, another potential monster fundraising for Anthropic.
The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.
All right, friends, quick announcements before we dive in.
First of all, thank you to today's sponsors, KPMG.
granola, robots and pencils, and Zencoder.
To get an ad-free version of the show, go to patreon.com slash ai daily brief, or you can subscribe on Apple podcasts.
If you want to learn more about sponsoring the show, send us a note at sponsors at ai daily brief dot AI.
Also, I just pushed at pulse dot ai daily brief dot AI, a new page to capture all of the learnings from our monthly pulse survey.
Check it out again at pulse dot ai daily brief dot AI.
We kick off today with more crazy numbers, things that seem normal now that would have been completely abnormal just like five minutes ago.
Anthropic is weighing up another round of fundraising at almost a trillion dollar valuation.
The Financial Times reports that Anthropic is considering one last round of private funding before their IPO, which is expected in the fall.
Anthropic last raised funds in February at a $380 billion valuation, but of course their revenue has gone parabolic since then.
And just to add a little heft to the market's appetite for this, There is a pre-IPO instrument that's trading on the blockchain on Jupiter, which is backed one-to-one by SPV exposure to Anthropic, basically sort of a crypto equivalent of Anthropic exposure, that about a week ago was trading at a $1.2 trillion implied valuation.
So at least some of the speculators out there think that this isn't crazy.
Now back in the mainstream world.
Sources told the FT that the round is expected to raise as much as $50 billion for Anthropic and which would value the company at $900 billion pre-money.
In the horse race, that would of course put Anthropic ahead of OpenAI, which last raised at $852 billion in March.
It also seems like demand is off the charts, with one potential investor commenting, people are ready to throw any dollar amount at Anthropic.
It's just about when Anthropic want to pop their heads up and say we're ready.
Another source noted that the SpaceX deal dramatically de-risked the investment, commenting, Anthropic has resolved the biggest bottleneck and potential source of weakness, which is compute.
Now, I don't know if I'd go so far as to say that that single deal resolves that, but it is notable that at least some investors are thinking that way.
Now, in terms of whether this comes together, Anthropic executives have reportedly held talks, but terms have not yet been agreed to.
And a last-minute fundraising round could be a big curveball for the twin AI IPOs later this year.
Anthropic establishing themselves as the top lab in valuation terms, at least, could help shift analysts' views.
It would also provide another private market valuation to support an IPO above a trillion dollars, instantly making Anthropic a top 20 public company.
Now, an AI-aligned IPO that is coming up much sooner is Cerebris.
Reuters reports that Cerebris is considering boosting their stock price ahead of this week's IPO.
The report says that the IPO price could be increased from a range of 115 to 125 per share to between 150 and 160 per share.
That increase would boost the implied valuation from the roughly $26 billion previously expected to more than $34 billion.
In addition, Cerebrus is considering upsizing the share offering from 28 million shares to 30 million shares.
All last week, reports claimed that demand was red-hot among institutional investors that were participating in the roadshow.
As of last Tuesday, demand was said to be roughly three times supply.
And Reuters now reports that Cerebrus has drawn orders for 20 times the number of available shares.
Trading is currently set to begin on Thursday in what will be the largest IPO so far this year.
Commentators are very mixed on this.
You can find plenty of folks who are bullish, but then others like Nvidia and tech analysts take him saying, I wouldn't touch Cerebrus with a 100-foot pole.
Sure, it may go up on hype sentiment in the short term, but it's far too fundamentally risky because of scaling and future execution.
Momar Chima points out that Polymarket is currently projecting Cerebrus to close above $50 billion in market cap by the end of day one.
One interesting note from TSMC, the company reported their slowest rate of sales growth in six months, despite the frenetic pace of the AI build-out.
The Taiwanese chipmaker reported 17.5% annualized sales growth for April, which is the slowest pace since last October.
And while 17.5% growth isn't slow per se, it was around half of what analysts had forecast.
Now, the question that people immediately raised to is whether this is a leading indicator of an AI bubble bursting.
Most, though, think that it has to either do with one, The fact that the company has experienced a huge slowdown in their non-AI business.
And two, just the sheer physical constraints of AI demand starting to eat into what TSMC can do.
Now, on the non-AI business side, the spiraling cost of memory, which is obviously due in large part to AI, has impacted consumer electronics and smartphone chip orders.
You combine that with global economic uncertainty and those segments plateaued over the first half of the year.
However, the biggest thing is it seems like TSMC just doesn't have any more fabs capable of manufacturing advanced AI chips.
In other words, they don't have excess capacity to fuel more growth right now.
And upstream components like high bandwidth memory are running into capacity constraints at the same time.
People were recently talking about a tweet from Intel CEO Lip Bhutan talking about collaborating to develop new products with NVIDIA.
And when someone said, looks like everyone is diversifying a little bit from TSMC?
Tay Kim again said, TSMC is sold out.
There's no choice.
Speaking of Intel, Apple has signed a preliminary chip-making agreement with the company, diversifying their own supply chain.
Until now, TSMC had been the sole producer of Apple's chips.
The Wall Street Journal reports that Intel will manufacture some of the chips for Apple devices, but it's unclear whether the deal is focused on lower-end iPhone and iPad chips or the top-of-the-line M-series.
The deal has been in negotiations for over a year, with the White House leaning on Apple hard to contract with Intel.
Following the government's investment last summer, Commerce Secretary Howard Lutnick has been meeting with tech companies including NVIDIA, Apple, SpaceX, and more to drum up business for Intel.
The journal noted that Apple was once TSMC's top customer, but skyrocketing demand for fab time has lessened their negotiating leverage in Taiwan.
The deal comes as Wall Street looks to a changing in leadership in semiconductors.
Mizuho analyst Jordan Klein said that there could be a, quote, changing of the guard in AI.
Both AMD and Intel have surged around 25% over the past week.
On the back of major deal-making activity, for example, last week AMD reached a deal with Rackspace.
Meanwhile, comparatively at least, Nvidia has struggled.
Their stock was up just 8% in a week where double-digit gains were the norm rather than the exception.
Mizuho's Klein noted that memory suppliers were the big winners, writing, That is what happens when a market quickly enters a material shortage condition and pricing surges higher while expenses rise only modestly.
You make a lot of money being overweight historic memory upturns when new capacity cannot be added fast enough.
It's that simple.
One wild little one staying on the chip theme.
Household data centers may be the next step to address the drastic compute shortage.
Major housing developers, including the Pulte Group, have entered a testing phase with partners including NVIDIA and California Startup Span to install micro data centers on newly built homes.
The units would be installed on housing exteriors and function as nodes in a distributed computing cluster.
Balaji Tamabatula, the COO of Baruban, which is a Texas energy and tech company, said, it's technically possible and already being explored.
He likened the plan to contribute and compute from an idle PC, just with enterprise-grade GPUs rather than consumer hardware.
He added feasibility depends on available power, internet connectivity, heat management, and the type of workload.
For batch processing and non-time-sensitive tasks, the home environment works surprisingly well.
Now, of course, the question will be whether homeowners and regulators agree to install networks of data centers throughout their communities.
Will the animosity towards large data centers come down in the same way?
It's not clear, but I think what this news suggests is that basically every single avenue is being explored right now to bring on more compute.
Lastly today, a new feature to play around with.
OpenAI has launched a new Chrome plugin for Codex to expand browser use.
Until now, Codex typically used an app-specific connector to access context from the browser.
This gave access to the web app, but meant Codex couldn't work alongside the user in a live session.
The new Chrome plugin gives Codex direct access to the browser.
It works in a separate tab group, meaning the user can keep working on their own session while Codex is running in the background.
The feature allows Codex to consider live browser context rather than just scraping data from a connector, which is particularly useful for web developers, where live browser context is often the only way to test functionality.
The feature, though, also could unlock powerful use cases for non-technical users.
The plugin allows Codex to quickly work across multiple tabs, navigating the web, filling in forms, and creating browser-native artifacts.
A small one on the face of it, but every time a new feature unlocks some specific set of use cases, it's worth taking note of.
For now, though, that is going to do it for today's AI Daily Brief Headlines edition.
Next up, the main episode.
One of the most important AI questions right now isn't who's using AI, it's who's using it well.
KPMG and the University of Texas at Austin just analyzed 1.4 million real workplace AI interactions and found something surprising.
The highest impact users aren't better prompt engineers, They treat AI like a reasoning partner.
They frame problems, guide thinking, iterate, and push for better answers.
And the good news?
These behaviors are teachable at scale.
If you're trying to move from AI access to real capability, KPMG's research on sophisticated AI collaboration is worth your time.
Learn more at kpmg.com slash US slash sophisticated.
That's kpmg.com slash US slash sophisticated.
Today's episode is brought to you by Granola.
Granola is the AI notepad for people in back-to-back meetings.
You've probably heard people raving about Granola.
It's just one of those products that people love to talk about.
I myself have been using Granola for well over a year now, and honestly, it's one of the tools that changed the way I work.
Granola takes meeting notes for you without any intrusive bots joining your calls.
During or after the call, you can chat with your notes, ask Granola to pull out action items, help you negotiate, write a follow-up email, or even coach you using recipes, which are pre-made prompts.
Once you try it on a first meeting, it's hard to go without.
Head to granola.ai slash ai daily and use code ai daily.
New users get 100% off for the first three months.
Again, that's granola.ai slash ai daily.
One thing I keep seeing in enterprise AI, companies hedging across every cloud, every model, every framework, or paying a GSI for a pilot that never ends.
The team's actually shipping, they've picked a lane, and they move fast.
That's one of the reasons I like today's sponsor, Robots and Pencils.
They've gone all in on AWS.
They're an advanced tier and AWS pattern partner, and they ship production AI co-workers in 45 days.
That's led to them doing some of the more interesting work I've seen on AI co-workers.
And by that, I'm not talking about chatbots.
I'm talking about actual agentic systems that sit inside a business architecture and do real work.
That kind of focus matters if you're an enterprise leader trying to get something real into production or an AWS rep trying to move a customer from interested to deployed.
Request an AI briefing at robotsandpencils.com.
One conversation with robots and pencils, and you'll know.
So coding agents are basically solved at this point.
They're incredible at writing code.
But here's the thing nobody talks about.
Coding is maybe a quarter of an engineer's actual day.
The rest is stand-ups, stakeholder updates, meeting prep, chasing context across six different tools.
And it's not just engineers.
Sales spends more time assembling proposals than selling.
Finance is manually chasing subscription requests.
Marketing finds out what shipped two weeks after it merged.
Zencoder just launched Zenflow Work.
It takes their orchestration engine, the same one already powering coding agents, and connects it to your daily tools.
Jira, Gmail, Google Docs, Linear, Calendar, Notion.
It runs goal-driven workflows that actually finish.
Your stand-up brief is written before you sit down.
Review cycle coming up?
It pulls six months of tickets and writes the prep doc.
Now you might be thinking, didn't OpenClaw try to do this?
It did, but it has come with a whole host of security and functional issues, which can take a huge amount of time to resolve.
Zencoder took a different approach.
SOC 2, Type 2 certified.
Curated integrations, tighter security perimeter, enterprise grade from day one, model agnostic and works from Slack or Telegram.
Try it at zenflow.free.
Welcome back to the AI Daily Brief.
One of the things that I like doing on this show is not just following the news, but following the conversation.
What I find is that what people are talking about tends to be not only the important slice of the news, as determined by what people are willing to spend their time and energy on.
But it's also the place where you find out how whatever's going on in the industry of AI is actually impacting what people are doing with AI.
And ultimately, I think most of you are here, at least in part, to understand what you can do with this technology.
And so when a post gets as much attention as this one from Tariq Shahipar from Anthropic, it's worth paying attention.
Now, what this post is nominally about is a proposed switch to how we interact and speak with agents.
We are, of course, now in the AI agent era.
Many, if not most of you at this point, based on our research, are using some combination of Claude Code or Codex or Cursor or something like that to build or manage agents that make your life and your work better.
And along with that, you've probably gotten pretty comfortable with the markdown format.
If you're anything like me, you have about a bajillion little MD files running around your computer in various places.
And in fact, that is at this point probably the primary way that you transfer big sets of information from one workspace to another.
For those of you who aren't using Markdown files or interacting with agents yet, but who do want to follow along with the conversation, an example I'll give you is that if I'm ever talking in Claude Online or ChatGPT Online, and I need to summarize the conversation to move it over into something like Claude Code or Codex, I'm going to have the web application write a handoff Markdown file that contains the summary and all the key details of the conversation we just had.
It is, in other words, a method of transferring context.
And yet Tariq, who works on Claude Code at Anthropic, and who has gotten pretty well known for his interesting technical essays, argued in a piece just before The Weeknd that became, in fact, The Weeknd's main topic of conversation, that we should think about ditching Markdown for HTML.
The piece, indeed, is called The Unreasonable Effectiveness of HTML, and at this point has been seen about 10 million times to give you a sense of just how engaged with this conversation people are.
Now, I'm going to read the highlights of the piece just to ground you in it before we get into the conversation and my argument for why Tariq is right, but for different reasons than he articulates.
My argument in short is that the reason that HTML is going to be in many cases a better format for transferring information than MD has to do with the very nature of the type of work we do and the fact that increasingly our jobs are shifting from producing a thing to giving an agent all that it needs to produce the thing on our behalf.
Coming back to Tariq's essay, he writes, Markdown has become the dominant file format used by agents to communicate with us.
It's simple, portable, has some rich text capability, and is easy for you to edit.
But as agents have become more and more powerful, I have felt that Markdown has become a restricting format.
I find it difficult to read a Markdown file of more than 100 lines.
I want richer visualizations, color and diagrams, and I want to be able to share them easily.
Now, a note from NLW here.
As you can see, a big part of Tariq's argument has to do with his interaction with Markdown.
Make a note of that as we proceed.
Tariq continues.
I'm also increasingly not editing these files myself, but using them as specs, reference files, brainstorming outputs, etc.
When I do make edits, I'm usually prompting Claude to edit them, which removes one of the Markdown's largest benefits.
I've started preferring HTML as an output format instead of Markdown, and increasingly see this being used by others on the Claude code team, and this is why.
Now, in the next section, he gets into the why of HTML, and gives a bunch of different reasons.
The first, he says, is information density.
HTML, he writes, can convey much richer information compared to Markdown.
It can, of course, do simple document structure like headers and formatting, but it can also represent all sorts of other information, such as tabular data using tables, design data with CSS, illustrations with SVG.
I would go so far as to say that there is almost no set of information that Claude can read that you cannot fairly efficiently represent with HTML.
This makes it a highly efficient way for the model to communicate in-depth information to you and for you to review with.
Another note from me here.
Keep in mind that question of efficiency, as that's going to be one of the major pushbacks on this piece.
The next in Tariq's list of why HTML is visual clarity and ease of reading.
And this is once again on the human side.
As Claude is able to do more complex work, he writes, it is also writing larger and larger specs and plans.
In practice, I found I tend to not actually read more than a 100-line markdown file, and I certainly am not able to get anyone else in my organization to read it.
But HTML documents are much easier to read.
Claude can organize the structure visually to be ideal to navigate with tabs, illustration, links, etc.
It can even be mobile responsive so you can read it differently based on your form factor.
The next benefit to HTML, he writes, is ease of sharing.
Markdown files are fairly hard to share since most browsers do not render them natively well.
You often have to add them as attachments to emails or messages.
With HTML, as long as you upload the file, you can share the link easily.
The chance of someone actually reading your spec, report, or PR write-up is much, much higher if it's in HTML.
The next benefit is what he calls two-way interaction.
HTML can allow you to interact with the document.
For example, you might want to ask it to add sliders or knobs or adjust to design or allow you to tweak different options in the algorithm to see what happens.
You can also ask it to let you copy these changes into a prompt to paste back into Claude code.
Lastly, he argues it's just more fun.
Making HTML documents with Claude is just more fun and makes me feel more involved and invested in the creation, and that by itself is enough.
Now moving into the practical, what you should do piece of the article is he points out, that this is not something that you necessarily need to add some new skill for, but which Claude, or of course he doesn't talk about ChatGPT, but ChatGPT can do well natively.
So what are some of the use cases he gives?
The first is specs, planning, and exploration.
HTML, he writes, is a rich canvas for Claude to dive into a problem.
When I start working on a problem instead of a simple markdown plan, I expect to make a web of HTML files.
For example, I might start with asking Claude code to brainstorm and create some explorations of different options.
I would then ask it to expand more into one or maybe make mockups or code snippets.
Finally, when I feel good, I'll ask it to write an implementation plan.
When I'm happy with the plan, I'll create a new session and pass on all of these files for it to implement.
Now, NLW inserting myself here again, make note of this use case, as this is where a big chunk of my argument will come in in just a few minutes.
Another use case he gives is code review and understanding.
He argues that code can be difficult to read in a markdown file, but with HTML, he can render diffs, annotations, flowcharts, modules, and more.
Design and prototypes, obviously this one is fairly self-explanatory.
Markdown is a text-only format.
HTML obviously has a lot more expressive options.
And then he also gives examples for reports, research, and learning, and custom editing interfaces.
So, like I said, this piece has about 10 million views.
And you have to think that as many crazy Claude Code and Codex users as there are out there, you only get to 10 million views if there's some amount of discussion around a piece, right?
Now, I will say that a fair number of the responses are just like this one from Daria and Nutmaz, who are adding their testimony to Tariq's idea.
Dario writes, I've now adapted this HTML approach and it has completely changed my vibe coding ability and quality.
So much better than hard to read Markdown files.
The spicier take, however, is summed up by Josh Dawes, who writes, The cynic in me can't help but note that HTML will cost way more tokens than Markdown.
And this was certainly one of the big strands of the conversation.
Now, it ran the gamut from fully cynical, Tariq is just a representative, if Anthropic, Anthropic's job is to sell more tokens, this sells more tokens.
to what I would say is the much more assuming of good faith, but still cautioning of if you go out and try this, you're probably going to increase your token usage, at least in a few cases where a markdown file would have worked just fine for much cheaper.
Some people shared their examples of how this had worked for them.
Jiayun-Ha was working on an explainer for the Codex slash goal feature, which we talked about a little bit on Friday's show, but which basically is Codex's version of a Ralph loop that works over and over again around a preset definitions of success.
allowing your coding agent to get a lot farther without you sitting there hovering and having to direct it.
Jiayun writes, started the codex slash goal explainer in Markdown last night, and about halfway through I was like, why am I fighting this?
Switch to HTML.
The flow diagram alone made it worth it, I can't really go back to be honest.
He shared the artifact that was produced, and so in this case for him, he was producing something public that he wanted to share, and Markdown was just ill-suited for that presentation format.
Specifically in this case I will note, because of his intended audience, that is a human consumer, validating kind of what Tariq said about the idea that he can't really read more than 100 lines of markdown and there's just some things that it doesn't make sense to visualize in text only.
Now, anytime we have a this versus that type of argument on the internet, it's a safe bet that the right answer is either A, probably somewhere in the middle, or B, a rejection of the premise that these things are actually diametrically opposed, and an appreciation for when and in what circumstances this is better versus that is better.
Josh Gonsalves sums that up in this particular argument.
HTML, he writes, is not a replacement for MD.
I have more MD files in my codebase than you have individual hairs on your head.
But I do agree that HTML is not a good replacement, but artifacts can actually be superior to Markdown in many cases, especially for interactive artifacts or visual ADHD builders like me.
So I think this is an and situation, not an or.
The smart ape wrote a follow-up post that does a pretty good job of trying to parse out when and in what circumstances one works better than another.
They write, Everyone is now picking sides.
Markdown fundamentalists who think HTML is showing off.
HTML converts who think Markdown is for people who haven't tried Cloud Code yet.
The question isn't Markdown versus HTML.
It's for this specific document, who reads it, who edits it, and how long does it live?
Answer those three and the format picks itself.
So going deeper, he says the three questions are audience, is it humans, Claude, in a future session or both?
Is the lifecycle written once or edited many times?
And is the horizon lives a day, a quarter, or forever?
In his argument, when it comes to audience, if Claude is reading it, that's a vote for markdown, whereas if humans are reading it, that's a vote for HTML.
On lifecycle, if it's going to be edited many times, that's a vote for markdown, whereas if it's going to be written once, that's a vote for HTML.
And on a horizon basis, if it's indexed and lasting, that's a vote for markdown.
Well, if it's ephemeral, that's a vote for HTML.
Now, of course, Ape says when the three votes line up, you have your answer, and when they split, you use a hybrid pattern.
So I think all of this is an interesting conversation, but what really fascinated me is that I think that this isn't just about ease of sharing and ease of human consumption, although those things are certainly true.
In my estimation, what's actually circling underneath this argument is a shift in what work means.
My argument is that in the agent era, the atomic unit of knowledge work has shifted.
The operator's job is in most cases not to finish the work, but instead to stage the conditions under which the agent can do the work.
Basically, we are moving from producing the final output to staging for the agent who produces the final output.
At various points in this conversation, I'll call it staging work or scaffolding work or priming work.
But you get the idea.
It's that work that needs to be done before the agent can really do its job well.
Now, I had already noticed this split a little bit in my own work.
And where it comes up most often is in handoffs between sessions.
It turns out that when you fully embrace the agent operator life, a lot of your time is spent on the setup.
I will frequently be shifting, for example, between the Cloud app and Cloud Code, or ChatGPT and Codex, or even between these different platforms, depending on whether I find one is better for a particular use case than the other.
And certainly the tool I use most often to share context from one to the other is an MD file, as I explained at the beginning of this conversation.
And yet a couple things started to happen.
First, I would very frequently need to also produce a separate HTML file for any sort of visual decision.
For example, mockups of a UI for a particular application, or even for just explaining more natively visual information like flowcharts or diagrams.
Now, to be clear, this is a workaround that I was using because of a capability deficit in the Markdown file itself.
The other thing that I would very frequently find myself doing is contextualizing on both the production end of the Markdown file and the consumption end of the Markdown file, adding a whole bunch of caveats.
For example, when I am handing off between a space in which I am brainstorming, and a workspace in which I'm going to be building, i.e.
from Claude app to Claude code, I always instruct the brainstorming space to try to make it clear what is and isn't actually decided based on our previous conversation.
The reason for that is that on average, when I ask an AI to write up a handoff document, it presents everything as though it has been ordained from on high.
It is not represented as what it actually is, which is a very in-between, in-medias-res sort of document.
and instead is the canonical summary of what is.
That constrains the building AI in ways that I have found not good.
Even when you're moving between Claude and Claude Code, or ChatGPT and Codex, the decisions that the builder agent is going to make are very frequently different than the suggestions of the brainstorming AI.
And so in addition to trying to build those caveats into the handoff markdown itself, I would also frequently find myself contextualizing for the receiving builder AI, Saying, don't take anything in here as gospel.
This is all still to be decided.
Ask questions here, debate with me, and we'll figure it out.
Point being that whether or not HTML is the answer, there are some big challenges that we're working through in this new type of in-between workspace.
And I think it's this in-between workspace, this liminal space, to use an overly academic but pretty accurate term, where we're just spending much more time than we ever used to.
Pre-AI, Your job as a knowledge worker was just to go from blank page to finished goal as fast as is reasonably possible.
The in-between states were all just transitional, to be exited onto the next thing as soon as you could.
In the AI, and specifically the agentic era, you are going to live in an in-between space where your job is to establish and maintain the conditions for finishing well for most of the process.
Even after you've handed off to the agent that is doing the production, you're going to stay in that liminal space Because when it produces something, you're going to need to go back in there to figure out what else needs to change.
This creates a calibration problem of the type that I just described.
If you've over-specified, you kill the agent's range and you take away the things it does better than you.
If you stay too vague, the agent can flail or produce generic output or ask endless clarifying questions.
The new skill is in calibrating how much structure to impose so that the unstructured remainder is something the agent can actually productively resolve.
And why this creates a challenge for markdown files specifically is that most times when you're living in this in-between liminal state, you're going to be operating in a state of mixed doneness.
What I mean is that some parts of the project are going to be locked, non-negotiable, decided.
Some areas, on the other hand, are going to be totally open, where the agent can explore, generate, or surface trade-offs.
And then, of course, some parts are going to be semi-decided, where you're leaning either a little or very hard into a particular decision.
But that decision is still provisional and subject to pressure testing.
And the point is that in most projects, a bunch of these different states are going to live next to each other at the same time.
The example that Claude came up with when I was describing this was a brief for building a new onboarding experience, where the goal and the platform were totally locked.
The goal was to reduce time to first value from 14 days to under 7.
The platform was web first, mobile in version 2, iOS native out of scope.
The visual system on the other end of the spectrum was totally open.
Maybe you want four different distinct directions, varying density and color confidence.
And then the flow user experience itself, as well as the tone, both had directions that the prompter was leaning in, but which were still provisional and at least somewhat open.
Now, of course, yes, a markdown file could take the words to describe those different states of doneness in the same document.
But to do this, you end up adding a lot of meta-commentary about the document inside the document.
A lot of things where you give it an instruction, but then you caveat that instruction in parentheses to follow.
And so HTML isn't without its challenges.
But the reason that I think it's interesting for some of these questions of the liminal state and mixed doneness is that you can use native features of HTML to encode that mixed doneness without having to get into that meta-commentary.
You can use tabs, progressive disclosure, side-by-side cards, annotation, expandable sections, visual hierarchy, color-coded status, interactive elements, all of which give the format itself the ability to communicate which parts are firm, which are exploratory, and which are open for the agent to figure out.
And even beyond the HTML versus markdown debate itself, I think that we are just at the beginning of exploring the right way to work in this new modality, where our job has shifted from the producers of the thing to the managers of the agent who produces the thing.
Indeed, you can kind of look at a lot of our operator conversations in that light.
When we talk about context engineering, that's just another version of this.
Where that question is, have we given the agent enough context for it to do its job well?
In some ways, this conversation that we've been having is a very particular slice of context, which is the context of the work process so far.
So that is our exploration for the day of the right way to talk to your agents.
Big thank you to Tariq for writing the article and getting this exploration rolling.
I'll be excited to see what people's individual experience with this is, but for today's AI Daily Brief, appreciate you listening or watching as always.
And until next time, peace.