What 73 builds actually taught me about working with agents

When I started, I expected to learn about AI. I did. I also expected to get faster at shipping. I did that too.

What I didn't expect: most of what I learned has nothing to do with AI specifically. It has to do with scope, with context, with knowing what you're actually building before you start.

What changed technically

Agents got significantly better at holding large codebases in context between 2024 and now. The earliest builds, I was constantly re-establishing what the project was at the start of every session. Now I drop in a context file — a few hundred words about the domain, the data model, the decisions I've already made — and the agent picks up almost exactly where we left off.

The MCP ecosystem matured in a way that genuinely changed my workflow. Tools that used to be reference material — Jira tickets, database schemas, meeting notes — are now live context that agents can query mid-conversation. That's not a small thing.

What didn't change

Judgment is still the bottleneck. Every build that went wrong went wrong because I handed off a decision I should have kept. Either I underspecified what "done" looked like, or I let the agent gold-plate something that should have shipped simple, or I didn't notice when the context had drifted from what I actually needed.

The first 10% and the last 10% are still yours. Agents are fast in the middle. Setting the direction at the start and doing the production hardening at the end — those still take the same amount of my attention they always did.

Real stakes produce better work. The builds I'm proudest of — the memorial site, the fundraising platform, the insurance knowledge base — all had a specific person who needed something real. The cleanest build I ever scoped was the one where I was building for a family in grief and every feature decision was obvious. Consequence focuses the mind.

The honest ledger

73 builds. 37 shipped. 19 useful tools I still reach for. 12 prototypes that taught me something. 5 experiments that mostly didn't work.

That's a good ratio. The 17 that didn't make it to "shipped" aren't failures — they're the tuition. I'd pay it again.