Building with AI: Substrate, Agents, Workflow

Creation is changing

Everything about creation is changing… but this isn’t some new thing caused by AI. Technology has been lowering barriers to creation and distribution for centuries. AI coding is just like the typewriter, or the high-level programming language, or the bedroom recording studio… New tools enable new creators to reach new heights.

But creating something great requires more than just tools:

• Taste: You have good instincts, prioritize correctly, etc.
• Skill: Your iterations always improve, and rarely regress.
• Experience: You have a lot of ideas and navigate them well.
• Energy: You iterate quickly and don't get tired or overwhelmed.

I've been honing these skills for two decades, creating music and building software. And I know from experience that the tool – the instrument, the software, the hardware – always matters far less than the medium and the message.

It always boils down to training, and then harnessing, all the taste, skill, experience, and energy of your ancestors, to create the best messages you possibly can, using whatever tools you have at your disposal.

Interestingly, this formula is also what agentic AI today boils down to (next token prediction with tool calling)…

Coding with AI in 2025

Everything is changing so quickly that I fully expect to be trying different tools every month for the foreseeable future, until the dust finally settles on AI coding, which could take years. And it will take far longer than that for the dust to settle on programming... Today, I use:

• Cursor to Explore and Plan
  • Using a slow & rigorous model (o3)
• Claude Code and Zed to Implement
  • Using a fast & eager model (Sonnet 4)

The AI coding system

Substrate, Agents, Workflow

When the dust settles on working with AI, we'll find that it all boils down to:

• The substrate, with a healthy immune system
• The agents, drivers for planning and doing
• The workflow, always deeply personal

This essay is mostly about coding with AI – but this framework applies to any creative pursuit, with or without AI.

With Zo Computer, we're building A General Interface for working with AI.

The AI coding iteration loop

The substrate, with a healthy immune system

Today, people generally report more success coding with AI on greenfield projects. This isn't just because greenfield projects are smaller. It's because projects that grow with AI from the outset naturally build a stronger immune system for receiving AI code.

Part of this AI immune system is old: comments, tests, compilers, linters, code review, observability, and continuous automation (with a blend of humans and machines in the loop). Just as mature codebases build stronger automated defenses as they deploy more frequently and add more contributors, healthy AI-first codebases must build strong automated defenses because the volume of output per individual contributor is unusually high.

The rest of the AI immune system is brand new. To build a healthy AI-first codebase in 2025, you’ll need:

• Rules that teach the AI to compile directions to code properly:
  • Cursor rules, CLAUDE.md, and the like are our new style guides and internal docs. Just like humans, AI agents only loosely follow these things. But they’re still important.
  • In addition to codebase-specific rules, I use these high-level system prompts
• An organization that teaches humans to direct AI properly:
  • Excelling with AI requires a mindset shift. It's not about being an individual contributor. It's not even about being a manager, or an architect, or a designer. It’s about being an organization. The great organizations of history have always been incredibly well-oiled machines: cyborg hive-minds, the perfect union of automation and precisely directed human activity.
  • The great organizations of the future will realize there’s a cyborg-hive-mind within each of us, waiting to be born.

The agents, drivers for planning and doing

Style becomes increasingly important as agents become more autonomous. It’s already happening, but soon aesthetic personality will become the primary factor when choosing an agent.

Today, agent choice is about optimizing for two distinct modes of work:

• Planning (what, why, how)

  • For this agent, I’m evaluating intelligence and depth. I’m looking at benchmarks like EnigmaEval and MultiChallenge. Here, the LLM matters more than the agent harness (I like Cursor, but any old harness will do). o3 is clearly the best model in my book: terse in writing style, exhaustive in context-seeking, and methodical in reasoning.

• Doing (compiling the plan into code)

  • For this, I’m evaluating speed and style. I like to route to different agents depending on the task:
  • Claude Code is my primary driver because it feels the most autonomous–capable of following long plans all the way to the end. But the coding style of Sonnet can be messy and overeager, and the lack of an editor can feel limiting. To deal with the lack of an editor, I review code being generated in Github Desktop, sometimes take a closer look or add commentary in Zed, and then do a final review in Graphite (I like their pull request UX, but I don’t believe in AI code review*)
  • I often switch back to Cursor and o3 for more complex implementation tasks–where I want it to feel more like pairing, and less like delegation.

*because false positive AI fatigue is the new alert fatigue.

The workflow, always deeply personal

You are not your tools. But you are the way you use your tools.

The great creators of history have always accomplished more than their peers using the same tools. Some people were just born to shred.

But many people learn to shred. You can learn to shred with AI, and you’ll eventually develop your own distinct style.

As you learn to build with AI, you’ll find the bottleneck is still your mortal human limits:

• Limited time
• Limited mental capacity

But with AI, you can expand your time. Here’s an easy way: work on two tasks in parallel. You can just use a single git branch, or try git worktree if you’re feeling fancy. It’s like braiding: just switch between two parallel chat threads until both tasks are done. If you select related tasks, you’ll be able to review both of them in a batch, and save additional context-switching overhead. Once you feel comfortable with two tasks, try braiding three…

Some say coding with AI leads to skill atrophy. This may be true: I’m noticeably worse now at typing code by hand. But I’ve also experienced something new and surprising: working with AI has expanded my mental capacity in strange and exciting ways.

Keep going. Collaborative creation with AI is going to be really great – I can already feel it.

The multi-agent AI coding workflow