Just like a sculptor needs to know how marble chisels, breaks, and buffs, a Web designer should know how CSS, HTML, and Javascript construct interfaces within a Web browser. You need to be intimate with your medium to know what it can and cannot do. Whether Web apps, iOS native apps, AI apps...

For years, many designers did exactly that. Looking at my personal GitHub history tells a story familiar to many. Steady coding until about 2014. Then almost nothing for a decade. Why?

Before 2014, a designer could build a lot with HTML, CSS, and Javascript. Then React and Angular gained traction, and "web app" went from "pages with some interactivity" to single-page applications with state management, routing, and build pipelines. The gap between "I can code a website" and "I can code in my team's dev environment" widened fast.

Tooling got heavier and frameworks churned constantly. Deployment went from dragging files to a server to CI/CD and cloud infrastructure. So little wonder that a designer who coded comfortably in 2012 could look at the 2015 landscape and reasonably decide to go back to Sketch (dated reference, I know.)

Thankfully technology never sits still and AI coding agents are now collapsing the gap between designing and building. Zooming in to the last few years of my GitHub history tells that story well.

For years, it was faster to mock up software than to ship it. Designers stayed "ahead" of engineering with prototypes. Now AI coding agents make development so much faster that the loop has flipped. Henry Modisett described this new state as "prototype to productize" rather than "design to build," and that sounds right to me.

Designers can now work iteratively with production code, not just prototypes. This kind of hands-on work creates better designers, ones who work through issues that previously got left for developers to figure out.

As always, designing software is better when you work in the medium, not a level or two abstracted away. AI tools make that possible again.

New Image Model

Google recently released a new version of their image generation model (Nano Banana 2) and I put it to the test on my Character Maker. The results are noticeably more dynamic and three-dimensional than the previous version. Characters have more depth, better lighting, and more active poses. So I'm now using it as the default model (until Reve 1.5 is available as an API).

One of the ways I originally reinforced consistency in my character maker was by checking whether an image generation model's API returned images with the same dimensions as the reference images I sent it. If the dimensions didn't match, I knew the model had ignored the visual reference so I forced it to try again. In my testing, this was needed about 1 in every 30-40 images. A very simple check, but it worked well.

A week into using Nano Banana 2, that sizing check started throwing errors. Generated images were no longer coming back with the exact dimensions of my reference images, breaking my verification loop. I had to resize the reference images to match Google's default 1K image size (1365px by 768px). But that took away my consistency check, so I had reinforce my prompt rewriter to make up for it.

Update: A day after publishing this overview, Google quietly changed the image format their API returns (from PNG to WEBP). This made image dimensions read incorrectly, causing every generation attempt to fail. Had to implementation a fix that works regardless of what format Google decides to send back.

Prompt Rewriter Iteration

This is where most of the ongoing work happens. As real people used the tool, edge cases piled up and the first step of my pipeline (prompt rewriting) had to evolve. For example, my character is supposed to be faceless (no eyes, no mouth, no hair). This had to be reinforced progressively over several iterations. Turns out image models really want to put a face on things.

For color accuracy, I shifted from named colors like "lime-green" that relied on the reference images for accuracy to explicitly adding both HEX codes and RGB values. Getting the exact greens to reproduce consistently required that level of specificity. I also added default outfit color rules for when people try to request color changes.

Content moderation expanded steadily as people found creative ways to push boundaries. I blocked categories like gore, inappropriate clothing, and full body color changes, while loosening rejection criteria from blocking any "appearance changes" to only rejecting clearly inappropriate inputs. The goal: allow creative freedom while preventing abuse.

The overall approach was: start broad, then iteratively tighten character consistency while expanding content moderation guardrails as real usage revealed what was needed.

At this point, my character comes back consistent almost every time. About 1 in 50 generations still produces an extra arm or a mouth (he's faceless, remember?). I've tested checking each image with a vision model then sending it back for regeneration if something is off (examples above). But given how rarely this happens and how much latency and cost it would auto check every image, it's currently not worth the tradeoff for me. For other uses cases, it might be?

If you haven't already, try the LukeW Character Maker yourself. Though I might have to revisit the pipeline again if you get too creative.

I opened by framing something I think is really important: every time there's a major technology platform shift, almost everything about what an "application" is changes. From mainframes to personal computers, from desktop software to web apps, from web to mobile, the way we build, deliver, and experience software transforms completely each time.

There's always this awkward period where we try to cram the old paradigm into the new one. I dug up an old deck from when we were redesigning Yahoo, and even two years after the iPhone launched, we were still just trying to port the Yahoo webpage into a native iOS app. The same thing is happening now with AI. The difference is this evolution is moving really, really fast.

From there, I walked through the stages of AI product evolution as I've experienced them.

The first stage is AI working behind the scenes. Back in 2016, Google Translate was "completely reinvented," but the interface itself changed not at all. What actually happened was they replaced all these separate translation systems with a single neural network that could translate between language pairs it was never explicitly trained on. YouTube made a similar move with deep learning for video recommendations. The UIs stayed the same; everything transformative was happening under the hood.

I remember being at Google for years where the conversation was always about how to make machine learning more of a core part of the experience, but it never really got to the point where people were explicitly interacting with an AI model.

That changed with the explosion of chat. ChatGPT and everything that looks exactly like it made direct conversation with AI models the dominant pattern, and chat got bolted onto nearly every software product in a very short time. I illustrated this with Ask LukeW, a system I built almost three years ago that lets people talk to my body of work in natural language. It seems pretty simple now, but building and testing it surfaced a few patterns that have carried over into everything we've done since.

One is suggested questions. When you ask something, the system shows follow-up suggestions tied to your question and the broader corpus. When we tested this, we found these did an enormous amount of heavy lifting. They helped people understand what the system could do and how to use it.

A huge percentage of all interactions kicked off from one of these suggestions. And they've only gotten better with stronger models. In our newer products like Rev (for creatives) and Intent (for developers), the suggestions have become so relevant that people often just pick them with keyboard shortcuts instead of typing anything at all.

Another pattern is citation. Even just seeing where information comes from gives people a real trust boost. In Ask LukeW, you could hover over a citation and it would take you to the specific part of a document or video. This was an early example, but as AI systems gain access to more tools and can do much more than look up information, the question of how to represent what they did and why in the interface becomes increasingly important.

And the third is what I call the walls of text problem. Because so much of this is built on large language models, people are often left staring at big blocks of text they have to parse and interpret. We found that bringing back multimedia, like responding with images alongside text, or using diagrams and interactive elements, helped a lot.

Through that walkthrough of what now seems like a pretty simple AI application, I'd actually touched on what I think are the three core issues that remain with us today: capability awareness (what can I do here?), context awareness (what is the system looking at?), and the walls of text problem (too much output to process).

The next major stage is things becoming agentic. When AI models can use tools, make plans, configure those tools, analyze results, think in between steps, and fire off more tools based on what they find, the complexity of what to show in the UI explodes. And this compounds when you remember that most of this is getting bolted into side panels of existing software. I showed a developer tool where a single request to an agent produced this enormous thread of tool calls, model responses, more tool calls, and on and on. It's just a lot to take in.

A common reaction is to just show less of it, collapse it, or hide it entirely. And some AI products do that. But what I've seen consistently is that users fall into two groups. One group really wants to see what the system is thinking and doing and why. The other group just wants to let it rip and see what comes out. I originally thought this was a new-versus-experienced user thing, but it honestly feels more like two distinct mindsets.

We've tried many different approaches. In Bench, a workspace for knowledge work, we showed all tool calls on the left, let you click into each one to see what it did, and expand the thinking steps between them. You could even open individual tool calls and see their internal steps. That was a lot.

As we iterated, we moved from highlighting every tool call to condensing them, surfacing just what they were doing, and eventually showing processes inline as single lines you could expand if you wanted. The pattern we've landed on in Intent is collapsed single-line entries for each action. If you really want to, you can pop one open and see what happened inside, but for the most part, collapsing these things (and even finding ways to collapse collapses of these things) is where we are now.

We also experimented with separating process from results entirely. In ChatDB, when you ask a question, the thinking steps appear on the left while results show up on the right. You can scroll through results independently while keeping the summary visible, or open up the thought process to see why it did what it did. Changing the layout to give actual results more prominence while still making the reasoning accessible has worked well.

On the capability awareness front, I showed several approaches we've explored. One is prompt enhancement, where you type something simple and the model rewrites it into a much more detailed, context-aware instruction. This gets really interesting when the system can automatically search a codebase (like our product Augment does) to find relevant patterns and write better instructions that account for them.

Another approach was Bench's visual task builder, where you compose compound sentences from columns of capabilities: "I want to... search... Notion for... a topic... and create a PowerPoint summarizing the findings." This gives people tremendous visibility into what the system can do while also helping them point it in the right direction.

And then there's onboarding. Designers are familiar with the empty screen problem, and the usual advice is to throw tooltips or tutorials at it. But it turns out we can have the AI model handle all of this instead. In ChatDB, when you drag a spreadsheet onto the page, the system picks a color, picks an icon, names the dashboard, starts running analysis, and generates charts for you. You learn what it does by watching it do things, rather than trying to figure out what you can tell it to do.

For context awareness, I showed how products like Reve let you spatially tell the model what to pay attention to. You can highlight an object in an image, drag in reference art, move elements around, and then apply all those changes. You're being very explicit through the interface about what the model should focus on. I also showed context panels where you can attach files, select text, or point the model at specific folders.

The final stage I explored is agents orchestrating other agents. In Intent, there's an agent orchestration mode where a coordinator agent figures out the plan, shows it to you for review, and then kicks off a bunch of sub-agents to execute different parts of the work in parallel. You can watch each agent working on its piece. I think there's a big open question here about where the line is.

How much can people actually process and manage? If you use the metaphor of being a manager or a CEO, can you be a CEO of CEOs? I don't think we know yet, but this is clearly where the evolution is heading.

The throughline of the whole talk was that while the final form of AI applications hasn't been figured out, certain patterns keep proving their value at each stage. Those durable patterns, the ones that hang around and sometimes become even more important as things evolve, are the ones worth paying close attention to.

The clearest example of this is in software development. In Anthropic's 2025 data, software developers made up 3% of U.S. workers but nearly 40% of all Claude conversations. A year later, their 2026 Measuring Agent Autonomy report showed software engineering accounting for roughly 50% of AI agent deployments. Whatever developers are doing with AI now, other domains are likely to follow suit.

And what developers have been doing is watching their role abstract upward at a pace that's hard to overstate.

• First, humans wrote code. You typed, the computer did what you said.
• Then machines started suggesting. GitHub Copilot's early form was essentially AI behind the scenes, offering inline completions. You picked which suggestions to use. Still very much in the driver's seat.
• Then humans started talking to AI directly. The chat era. You could describe what you wanted in natural language, paste in a broken function, brainstorm architecture. The model became a collaborator.
• Then agents got tools. The model doesn't just respond with text anymore. It searches files, calls APIs, writes code, checks its own work, and decides what to do next based on the results. You're no longer directing each step.
• Then came orchestration. A coordinator agent receives your request, builds a plan, and delegates to specialized sub-agents. You review and approve the plan, but execution fans out across multiple autonomous workers.

To make this more tangible, our developer workspace, Intent, makes use of agent orchestration where a coordinator agent analyzes what needs to happen, searches across relevant resources, and generates a plan. Once you approve that plan, the coordinator kicks off specialized agents to do the work: one handling the design system, another building out navigation, another coordinating their outputs. Your role is to review, approve, and steer.

Stack that one more level and you've got machines running machines running machines. At which point: where exactly does the human sit?

To use a metaphor we're all familiar with: a manager keeps tabs on a handful of direct reports. A director manages managers. A CEO manages directors. At each layer, the person at the top trades direct understanding for leverage. They see less of the actual work and more of the summaries, status updates, and roll-ups.

But being an effective CEO is extraordinarily rare. Not just thinking you can do it, but actually doing it well. And a CEO of CEOs? The number of people who have operated at that scale is vanishingly small.

Which raises two questions. First, how far up the stack can humans actually go? Agent orchestration? Orchestration of orchestration? Where does it break down? Second, at whatever level we land on, what skills do people need to operate there?

The durable skills may turn out to be steering, delegation, and awareness: knowing what to ask for, how much autonomy to grant, and when to look under the hood. These aren't programming skills. They're closer to the skills of a good leader who knows when to let the team run and when to step in.

We used to design how people interact with software. Now we're designing how much they need to.

When Twitter acquired Bagcheck in 2011, Sam (CTO) and I were shipping multiple times a day. We started with a command line interface that let us figure out what objects and actions we needed before ever building any UI. When we did, we used logic-less templates so I could iterate on the front-end quickly while Sam managed the back-end code.

The point was to move fast and learn. With just two people building the product, we never got bottlenecked on decision-making or coordination. While conventional wisdom says "add more resources" to go faster, it rarely works out that way. Most companies go slow because of plodding decision making and opaque alignment. Smaller teams naturally don't have this problem.

But small teams can only do so much right? That's why every team in a big company is always asking for more resources. Not anymore.

Armed with highly capable AI systems, everyone (designer, developer, etc.) on a team can get more done. In big teams, though, these new capabilities smack head first into the decision-making and alignment problems that have always been there. In small teams, they don't.

So how small? Surely we need at least 100? 50? Bagcheck never crossed four employees and when Google acquired my next company, Polar, in 2014 there was five of us. These companies pre-dated AI coding agents and large language models. With today's AI capabilities, the number of people you need to get a lot done fast is probably a lot smaller than you think.

An AI product becomes agentic when the model doesn't just respond to a prompt, but plans which tools to use, configures them, and decides its next steps based on the results. This additional set of process means AI products are able to do more, check their work, and thereby provide better results. The downside, though, is it can be a lot for people to take in.

Whether people are using agentic products for coding, data analysis, or writing, I keep seeing the same split: some users find the agent's work overwhelming and want the interface to focus purely on results. Others say seeing that work is essential for monitoring and checking what the agent is doing. Strongly worded feedback comes in from both sides.

I initially assumed this was a temporary divide. New users tend to watch closely and check the system's progress, but as trust builds, that scrutiny fades and monitoring starts to feel like a chore. Yet it still seems like there's two camps (for now). So how does a product strike the balance?

When working on Bench, a workspace for knowledge work, we explored many approaches to displaying tool use, results, and configuration. (though we quickly learned, no one configured tools, that's the agent's job.) In this exploration, results from each tool are grouped beneath it and open in the right column when selected (video below).

A later iteration featured several levels of progressive disclosure. Tool calls were collapsed by default, and selecting one would show its results in the right column. Selecting the timeline highlighted all the process and decision points between tool uses. You could even open each tool's settings, re-run it, or stop it mid-execution (video below). Tools were new back then and we were working off the assumption that people would want visibility and control. It was too much.

In subsequent iterations we focused on reducing the visual weight of tools and showing less process by default. This became even more important as the number of tools grew..

For ChatDB, which helps people understand and visualize data, we split the interface into two columns. While the agent works (video below), the left side shows what it's doing: the decisions it's making, the tools it's picking, and so on. When results appear in the right column, the left side collapses down to a summary and link so the focus shifts to the output. Anyone who wants to review the steps can open it back up.

This approach allows the agent's work to serve a detailed progress indicator, instead of forcing people to watch a spinner while things work.

More recently in Intent, a developer workspace for working with agents, we used a single line to show an agent's work with the ability to expand it for more details. It's an attempt to strike a balance between too much and not enough but I still hear opinions on both sides.

I charted the transition from AI models behind the scenes to chat to agents last year in The Evolution of AI Products. At the time, we were wrestling with how to spin up sub-agents and run them in the background. That's mostly been settled and agent orchestration (coordinating and verifying the work of multiple agents on unified tasks) is today's AI product design challenge.

As Microsoft CEO, Satya Nadella put it:

"One of the metaphors I think we're all sort of working towards is 'I do this macro delegation and micro steering [of AI agents]'. What is the UI that meets this new intelligence capability? It's just a different way than the chat interface. And I think that would be a new way for the human computer interface. Quite frankly, it's probably bigger."

He's right. When you have multiple agents working together, you need more than a conversation thread as anyone that's tried to manage a team through a single Slack or email thread can attest.

Introducing Intent

Intent by Augment (in early preview today) is a new software development app with agent orchestration at its core. You're not managing individual model calls or chat threads. You're setting up workspaces, defining your intent (what you want to get done), and letting specialized agents work in parallel while staying aligned.

To ground this in a real-world analogy, if you want to accomplish a large or complicated task you need...

• A team of the right people for the job, often specialists
• To give the team the information they need to complete the job
• The right environment where the team can coordinate and work safely

That's a space in Intent in a nutshell. Software developers create a new space for every task they want to get done. Each space makes use of specific agents and context to complete the task. Each space is isolated using git worktrees so agents can work freely and safely. Fire up as many spaces as you want without having them interfere with each other.

I've often said "context is king" when talking about what makes AI products effective. That's especially true when you need to coordinate the work of multiple parallel agents with varying capabilities. In Intent, context is managed by a living spec which provides a shared understanding that multiple agents can reference while working on different parts of a problem. This living spec is written and updated by a coordinator agent as it manages the work of implementer and verifier agents. It's a whole agent dev team.

Because agents operate from the same spec, parallel work becomes possible. Assumptions, tradeoffs, and decisions stay aligned and updated as code changes without requiring constant human intervention to keep things on the same page. For instance, one agent handles the theme system while another works on component styles. Both reference the same context, so their work fits together.

By default, a coordinator writes a spec and delegates to specialists for you. But you can also set up spaces with custom agents and manage your own context if you want. Think of it as manual vs. auto mode.

The UI for agent orchestration in Intent isn't a fancier chat interface. It's context management, agent specialization, and a unified developer workflow. It's not hard to squint and see very similar orchestration UI being useful for lots of other domains too.

For decades, designers delivered fixed assets. A project meant a set number of ads, illustrations, mockups, icons. When the client needed more, they came back to the designer and waited. To help others create on-brand assets without that bottleneck, designers crafted brand guidelines: documents that spelled out what could and couldn't be done with colors, typography, imagery, and layout.

But with today's AI coding agents, building software is remarkably easy. So instead of handing over static assets and static guidelines, designers can deliver custom software. Tools that let clients create their own on-brand assets whenever they need them.

This is something I've wanted to build ever since I started using AI image generators within Google years ago. I tried: LoRAs, ControlNet, IP-Adapter, character sheets. None of it worked well enough to consistently render assets the right way. Until now.

LukeW Character Maker

Since the late nineties, I've used a green avatar to represent the LukeW brand: big green head, green shirt, green trousers, and a distinct flat yet slightly rendered style. So to illustrate the idea of design tools as deliverables, I build a site that creates on-brand variations of this character.

The LukeW Character Maker allows people to create custom LukeW characters while enforcing brand guidelines: specific colors, illustration style, format, and guardrails on what can and can't be generated. Have fun trying it yourself.

How It Works

Since most people will ask... a few words on how it works. A highly capable image model is critical. I've had good results using both Reve and Google's Nano Banana but there's more to it than just picking an image model.

People's asset creation requests are analyzed and rewritten by a large language model that makes sure the request aligns with brand style and guidelines. Each generation also includes multiple reference images as context to keep things on rails. And last but least, there's a verification step that checks results and fixes things when necessary. For instance, Google's image generation API ignores reference images about 10-20% of the time. The validation step checks when that's happening and re-renders images when needed. Oh, and I built and integrated the software using Augment Code.

The LukeW Character Maker is a small (but for me, exciting) example of what design deliverables can be today. Not just guidelines. Not just assets. But Tools.

Reve's recent launch of Effects illustrates this transition. Want a specific film grain look for your image or video? Just describe it in plain language or upload an example. Reve's AI agent will write code that produces the effect you want and figure out what parameters should be adjustable. Those parameters then become sliders in an interface built for you in real-time.

Instead of having to find the menu item for an existing filter (if it even exists) in traditional software, you just say what you want and the system constructs it right then and there.

When applications can generate capabilities on demand, the definition of "what this product does" becomes more fluid. Features aren't just what shipped in the last release, they're also what users will ask for in the next session. The application becomes a platform for creating its own abilities, guided by user intent rather than predetermined roadmaps.

In Context Management UI in AI Products I looked at UI patterns for showing users what information is influencing AI model responses, from simple context chips to nested agent timelines. This time I want to highlight two examples of automatic and manual context management solutions.

Augment Code's Context Engine demonstrates how automatic context management can dramatically improve AI product outcomes. Their system continuously indexes code commit history (understanding why changes were made), team coding patterns, documentation, and what developers on a team are actively working on.

When a developer asks to "add logging to payment requests," the system identifies exactly which files and patterns are relevant. This means developers don't have to manually specify what the AI should pay attention to. The system figures it out automatically and delivers much higher quality output as a result (see chart below).

Having an intelligent system manage context for you is extremely helpful but not always possible. In many kinds of tasks, there is no clear record of history, current state, and relevance like there is in a company's codebase. Also, some tasks are bespoke or idiosyncratic meaning only the person running them knows what's truly relevant. For these reasons, AI products also need context management interfaces.

Reve's creative tooling interface not only makes manual context management possible but also provides a consistent way to reference context in instructions as well. When someone adds a file to Reve, a thumbnail of it appears in the instruction field with a numbered reference. People can then use this number when writing out instructions like "put these tires @1 on on my truck @2".

It's also worth noting that any file uploaded to or created by Reve can be put into context with a simple "one-click" action. Just select any image and it will appear in the instruction field with a reference number. Select it again to remove it from context just as easily.

While the later may seem like a clear UI requirement, it's surprising how many AI products don't support this behavior. For instance, Google's Gemini has a nice overview panel of files uploaded to and created in a session but doesn't make them selectable as context.

As usual, AI capabilities keep changing fast. So context management solutions, whether automatic or manual, and their interfaces are going to continue to evolve.

As AI coding agents continue to improve dramatically, developers are turning to them more and more to not only write code but to review and improve it as well. The result isn't just more coder faster but the organizational changes needed to support this transition as well.

"The vast majority of code that is used to support Claude and to design the next Claude is now written by Claude. It's just the vast majority of it within Anthropic. And other fast moving companies, the same is true."
- Dario Amodei, Anthropic CEO

"Codex has transformed how OpenAI builds over the last few months."
- Sam Altman, OpenAI CEO

As just one example, a product manager I speak with regularly now spends his time using Augment Code on his company's production codebase. He creates a branch, prompts Augment's agents until he has a build he's happy with then passes it on to Engineering for implementation. Instead of writing a Product Requirements Document (PRD) he creates code that can be used and experienced by the whole team leading to a clearer understanding of what to build and why.

This kind of accelerated prototyping is a common way for designers to start applying AI coding agents to their workflow as well. But while the tools may be new, prototyping isn't new to designers. In fact, many larger design teams have specific prototyping roles within them. So what additional capabilities do AI coding agents give designers? Here's a few I've been using regularly.

Note: It's worth calling out that for these use cases to work well, you need AI coding tools that deeply understand your company's codebase. I, like the PM mentioned earlier, use Augment Code because their Context Engine is optimized for the kinds of large and complex codebases you'll find in most companies.

Fix Production Bugs

See a bug or user experience issue in production? Just prompt the agent with a description of the issue, test its solution, and push a fix. Not only will fixing bugs make you feel great, your engineering friends will appreciate the help. There's always lots of "small" issues that designers know can be improved but can't get development resources for. Now those resources come in the form of AI coding agents.

Learn & Rethink Solutions

Sometimes what seems like a small fix or improvement is just the tip of an iceberg. That is, changing something in the product has a fan-out effect. To change this, you also need to change that. That change will also impact these things. And so on.

Watching an AI coding agent go through its thinking process and steps can make all this clear. Even if you don't end up using any of the code it writes, seeing an agent's process teaches you a lot about how a system works. I've ended up rethinking my approach, considering different options and ultimately getting to a better solution than I started with. Thanks AI.

Get Engineering Involved

Prompting an agent and seeing its process can also make something else clear: it's time to get Engineering involved. When it's obvious the scope of what an AI agent is trying to do to solve an issue or make an improvement is too broad, chances are it's time to sit down with the developers on your team to come up with a plan. This doesn't mean the agent failed, it means it prompted you to collaborate with your team.

Through these use cases, AI coding agents have helped me make more improvements and make more informed improvements to the products I work on. It's a great time to be a designer.

Today's AI chat interfaces hit usability issues when models engage in extended reasoning and tool use (aka they get more agentic). Instead of simple back-and-forth chat, these conversations look more like long internal monologues filled with thinking traces, tool calls, and multiple responses. This creates UI problems, especially in narrow side panels where people lose context as their initial instructions and subsequent steps are off-screen while the AI continues to work and evaluate its results.

As you can see in the video above, our dual-scroll pane layout addresses these issues by separating an AI model's process and results into two columns. User instructions, thinking traces, and tool calls appear in the left column, while outputs show up in the right column.

Once the AI completes its work, the thinking steps (traces, tool calls) collapse into a summary on the left while results remain persistent and scrollable on the right. This design keeps both instructions and outcomes visible simultaneously even when people move between different instructions. Once done, the collapsed thinking steps can also be re-opened if someone needs to review an AI model's work. Each step in this process list is also a link to its specific result making understanding and checking an AI model's work easier.

You can try out these interactions yourself on ChatDB with an example like this retail site or with your own data.

Thanks to Sam Breed and Alex Godfrey for the continued evolution of this UI.

Chat is a pretty simple and widely understood interface pattern... so what's the problem? Well when it's just two people talking in a messaging app, things are easy. But when an AI model is on the other side of the conversation and it's full of reasoning traces and tool calls (aka it's agentic), chat isn't so simple anymore.

Instead of "you ask something and the AI model responds", the patterns looks more like:

• You ask something
• The model responds with it's thinking
• It calls a tool and shows you the outcome
• It tells you what it thinks about the outcome
• It calls another tool ...

While these kinds of agentic loops dramatically increase the capabilities of AI models, they look a lot more like a long internal monologue than a back and forth conversation between two people. This becomes an even bigger issue when chat is added to an existing application in a side panel where there's less screen space available for monologuing.

Using Augment Code in an development application, like VS Code, illustrates the issue. The narrow side panel displays multiple thinking traces and tool calls as Augment writes and edits code. The work it's doing is awesome, staying on top of it in a narrow panel is not. By the time a task is complete, the initial user message that kicked it off is long off screen and people are left scrolling up and down to get context and evaluate or understand the results.

That this point design teams start trying to sort out how much of the model's internal monologue needs to be shown in the UI or can parts of it be removed or collapsed? You'll find different answers when looking at different apps. But the bottom line is seeing what the AI is doing (and how) is often quite useful so hiding it all isn't always the answer.

What if we could separate out the process (thinking traces, tool calls) AI models use to do something from their final results? This is effectively the essence of the chat + canvas design pattern. The process lives in one place and the results live somewhere else. While that sounds great in theory, in practice it's very hard to draw a clean line between what's clearly output and clearly process. How "final" does the output need to be before it's considered "the result"? What about follow-on questions? Intermediate steps?

Even if you could separate process and results cleanly, you'd end up with just that: the process visually separated from the results. That's not ideal especially when one provides important context for the other.

To account for all this and more, we've been exploring a new layout for AI chat interfaces with two scroll panes. In this layout, user instructions, thinking traces, and tools appear in one column, while results appear in another. Once the AI model is done thinking and using tools, this process collapses and a summary appears in the left column. The results stay persistent but scrollable in the right column.

To illustrate the difference, here's the previous agentic chat interface in ChatDB (video below). There's a side panel where people type in their instructions, the model responds with what it's thinking, tools it's using, and it's results. Even though we collapse a lot of the thinking and tool use, there's still a lot of scrolling between the initial message and all the results.

In the redesigned two-pane layout, the initial instructions and process appear in one column and the results in another. This allows people to keep both in context. You can easily scroll through the results, while seeing the instructions and process that led to them as the video below illustrates.

Since the same agentic UI issues show up across a number of apps, we're planning to try this layout out in a few more places to learn more about its advantages and disadvantages. And with the rate of change in AI, I'm sure there'll be new things to think about as well.

AI broke our networking assumptions. Traditional networking expected some latency variance and occasional failures. AI workloads demand perfection: high bandwidth, ultra-low jitter (tens of microseconds), and near-flawless reliability. One slow node kills the entire training job.

Why AI is different: These workloads use bulk synchronous parallel computing. Everyone waits at a barrier until every node completes its step. The slowest worker determines overall speed. No "good enough" when 99 of 100 nodes finish fast.

Real example: Gemini traffic shows hundreds of milliseconds at line rate, but average utilization is 5x below peak. Synchronized bursts with no statistical multiplexing benefits. Both latency sensitive AND bandwidth intensive.

Three Breakthroughs

Falcon (Hardware Transport): Existing hardware transports assumed lossless networks: fundamentally incompatible with Ethernet. Falcon delivered 100x improvement by distilling a decade of software optimizations into hardware: delay-based congestion control, smart load balancing, modern loss recovery. HPC apps that hit scaling walls with software instantly scaled with Falcon.

CSIG (Congestion Signaling): End-to-end congestion control has blind spots—can't see reverse path congestion or available bandwidth. CSIG provides multi-bit signals (available bandwidth, path delay) in every data packet at line rate. No probing needed. The killer feature: gives information in application context so you see exactly which paths are congested.

Firefly: Jitter kills AI workloads. Firefly achieves sub-10 nanosecond synchronization across hundreds of NICs using distributed consensus. Measured reality: ±5 nanoseconds via oscilloscope. Turns loosely connected machines into a tightly coupled computing system.

The Remaining Challenges

Straggler detection: Even with perfect networking, finding the one slow GPU in thousands remains the hardest problem. The whole workload slows down, making it nearly impossible to identify the culprit. Statistical outlier analysis is too noisy. Active work in progress.

Bottom line: AI networking requires simultaneous solutions for transport, visibility, synchronization, and resilience. Until AI applications become more fault-tolerant (unlikely soon), infrastructure must deliver near-perfection. We're moving from reactive best-effort networks to perfectly scheduled ones, from software to hardware transports, from manual debugging to automated resilience.

In all the technology companies I've worked at, big and small, there's always been a mindset of "we don't have enough resources to get everything we want done." Whether that's an excuse or not, companies consistently strive for more productivity. Well, now we have it.

More and more developers are finding that today's AI coding agents massively increase their productivity. As an example, Amazon's Joe Magerramov recently outlined how his "team's 10x throughput increase isn't theoretical, it's measurable." And before you think "vibe coding, crap" his post is a great walkthrough on how developers moving at 200 mph are cognizant of the need to keep quality high and rethink a lot of their process to effectively implement 100 commits a day vs. 10.

But what happens to software design teams when their development counterparts are shipping 10x faster? I've seen three recurring reactions:

1. Our Role Has Changed
2. We're Also Faster Now
3. It's Just Faster Slop

Our Role Has Changed

Instead of spending most of their time creating mockups that engineers will later be asked to build, designers increasingly focus on UX alignment after things are built. That is, ensuring the increased volume of features developers are coding fit into a cohesive product experience. This flips the role of designers and developers.

For years, design teams operated "out ahead" of engineering, unburdened by technical debt and infrastructure limitations. Designers would spend time in mockups and prototypes envisioning what could be build before development started. Then developers would need to "clean up" by working out all the edge cases, states, technical issues, etc that came up when it came time to implement.

Now development teams are "out ahead" of design, with new features becoming code at a furious pace. UX refinement and thoughtful integration into the structure and purpose of a product is the "clean up" needed afterward.

We're Also Faster Now

An increasing number of designers are picking up AI coding tools themselves to prototype and even ship features. If developers can move this fast with AI, why can't designers? This lets them stay closer to the actual product rather than working in abstract mockups. At Perplexity, designers and engineers collaborate directly on prompting as a programming language. At Sigma, designers are fixing UX issues in production using tools like Augment Code.

It's Just Faster Slop

The third response I hear is more skeptical: just because AI makes developers faster doesn't mean it makes good products. While it feels good to take the high ground, the reality is software development is changing. Developers won't be going back to 1x productivity any time soon.

"Ninety percent of everything is crap" - Sturgeon's law

It's also worth remembering Sturgeon's Law which originated when the science fiction writer was asked why 90% of science fiction writing is crap. He replied that 90% of everything is crap.

So is a lot of AI-generated code not great? Sure, but a lot of code is not great period. As always it's very hard to make something good, regardless of the tools one uses. For both designers and developers, the tools change but the fundamental job doesn't.