Amelia Wattenberger: Designing The Next Flow State
TL;DR
Amelia Wattenberger thinks developers are grieving a lost flow state — she says the old rhythm of mastery-through-syntax is breaking as agents take over implementation, and the real challenge now is building tools that restore flow in a more intention-heavy way.
The spec is becoming the new source of truth — in Intent, Wattenberger argues teams need a living document that sits between a three-word prompt like “add dark mode” and the full codebase, with humans and agents both updating it as work evolves.
She frames AI tooling as a climb up abstraction layers, but says purpose is where automation stops — from punch cards to TypeScript to agents and “orchestrating the orchestrators,” her memorable line is that agents won’t eat ice cream for her because the point is the experience itself.
AI coding tools have moved through distinct eras already — she maps the progression from GitHub Copilot’s “fancy autocomplete,” to chat in the IDE, to agentic IDEs, to the CLI/tmux phase, and now toward ADEs like Intent that bring interfaces back for visibility and control.
Her core design problem is compression, not raw information — drawing on her data-vis background, she says agent workflows need lossy summaries that show the gist without forcing people to read every chat log, while still letting them pop the hood and inspect the engine.
Intent’s workspace model is her answer to the chaos of juggling many AI tasks — instead of centering the chat thread, she organizes work into isolated task bundles that behave like “20 desks,” each preserving specs, agents, previews, and progress so you can context-switch without losing the plot.
The Breakdown
A career built by moving up and down the stack
Tommy opens by framing AI as one of the wildest moments in humanity, full of both hope and labor-market anxiety, then positions Amelia Wattenberger as someone unusually equipped to think through it. She describes her own path from eight years as a front-end developer at small analytics startups into design-heavy roles, saying code, pixels, and Figma are all just materials for the same thing: turning an idea in her head into something real.
From buttons and components to agents and abstraction
Amelia explains her long-running instinct as moving from vertical slices like HTML/CSS/JS toward horizontal slices around the thing itself — “it’s a button and I want to work on the button.” She sees AI as the next abstraction jump after higher-level programming languages: more leverage, more distance from the raw material, and a weird infinite ladder of “just throw an agent at it” until you hit actual human purpose.
“Agents aren’t going to eat ice cream for me”
That’s her cleanest line in the interview, and it lands because it answers where abstraction stops. Tommy compares the industry to a Rube Goldberg machine of ever-more plumbing, and Amelia pushes back by saying the end state isn’t infinite automation — it’s whatever part of the work is meaningful in the first place.
Why developers feel off-balance right now
Her most resonant point is that developers are mourning the old flow state: the satisfying, syntax-driven groove where mastery of tools made the work feel smooth and earned. Agents break that cadence — you ask for something, wait 20 minutes, start something else, and wind up in a “herky-jerky” workflow — so she thinks the next generation of tools has to bring flow back around planning, shaping, and intent rather than raw implementation.
The workflow shifts from code to medium-switching
Amelia describes building as a loop of plan, implement, review that exists both at macro and micro levels, but says AI is squeezing most of the human effort into the front end of that loop. Her key idea is that good work now moves through multiple media — natural language, specs, diagrams, Figma, prototypes, code — each with different “physics” that help tease out assumptions at the right fidelity.
The eras of AI coding tools, from Copilot to the ADE
She gives a great mini-history: GitHub Copilot as “fancy autocomplete,” then chat in the IDE, then agentic chat that can write code, then the CLI/tmux backlash where people tore down the IDE to rethink workflows. Now, she says, the pendulum is swinging back because interfaces are actually useful: Intent is an “ADE,” an AI-native developer environment built around visibility, observability, and recursive drill-down instead of chat logs and tiny terminal titles.
Compression is the real interface problem
Drawing from data visualization, Amelia says the challenge isn’t exposing everything agents are doing — it’s compressing the signal so humans can focus. Like a bar chart that throws away texture to reveal a pattern, a good AI interface should summarize what matters, let you stay oriented, and still allow you to zoom from a high-level map down to the IKEA-shelf level when something needs inspection.
Intent, living specs, and the new workspace
When Tommy asks where a team’s source of truth lives, Amelia says it can’t be buried in chat logs any more than a codebase can be understood from Git commits alone. Intent’s answer is a workspace-centered model with notes, subtasks, and a “living spec” that both humans and agents update; she demos it like a room with many desks, each task isolated but resumable, whether it’s a blog post, dark mode feature, PR review, or speculative whiteboarding experiment.
How she actually thinks and builds in this moment
The conversation ends with her creative process: long walks, ChatGPT as a sounding board, sketches on paper, Figma, and narrative blog posts like “Our interfaces have lost their senses.” Her advice to busy engineers and designers is simple but urgent: keep playing with new tools because their physics are changing fast, and even a workflow that crashes and burns or costs $10 teaches you more than sitting still while the stack shifts under you.