AI-First Development: How I Work With Agents

Not "AI wrote my app". Agents run my audits, refactors and releases, but only because I know what should come out. What I delegate, what I never hand off.

"AI-First" Is Not Autocomplete

It's not editor suggestions. It means the default executor of a task is an agent, and I design the task, supply the context, and review the work like I would from a teammate.

In my earlier post about cmux I described the layout: every project gets a workspace, and the first pane of each one is an agent. That's not decoration. Most of my work flows through that pane.

This Week's Example: SEO From Audit to Deploy

I asked an agent (Claude Code) to audit this blog's indexing. In a single session the agent:

• found why the site wasn't indexed: every subpage declared the homepage as its canonical URL, so Google treated the whole site as one page,
• rewrote the build script to generate per-page meta tags and canonicals, plus a sitemap with real dates,
• deployed the changes, verified them in production, and walked me through Google Search Console verification.

My role: the assignment, two decisions along the way, and a diff review. Minutes, not hours. That's the leverage "AI-first" is about.

What I Delegate, and What I Don't

I delegate almost always: "why doesn't this work" diagnoses, because an agent reads more logs and code in a minute than I do in an hour (that's how the hardcoded version in Info.plist was found, which I wrote about here); refactors with a clearly defined end state; research with sources required; boilerplate and unit tests for existing logic.

I don't delegate: product decisions and architecture, because an agent executes a vision brilliantly, but it still has to be someone's vision; judging whether code is production-ready, because I read the diff myself, always; conversations with humans, because automating those is a false economy.

Three Rules That Make the Difference

1. Context is currency. An agent with access to the repo, the tickets and the project's conventions produces senior-level output. An agent with a bare prompt produces Stack Overflow level. I maintain convention files precisely so every session starts fully informed.
2. Delegate outcomes, not steps. "Fix the blog's indexing" works better than dictating one command at a time. Micro-managing an agent costs more than doing the thing yourself.
3. Verification is still engineering. An agent can build an almost-right thing with full confidence. The gap between "almost" and "right" is my responsibility: tests, diffs, actually running the app. AI-first doesn't mean AI-only.

What It Actually Buys

I won't quote you "10x productivity" because I don't measure myself in marketing units. What I observe: I run, in parallel, work that used to require choosing "this or that": a commercial product, two products of my own, marketing, research.

That's probably the most honest definition of AI-first in 2026: not writing code faster, but raising the throughput of the whole person. The code was never the bottleneck anyway.