Building With Agentic Workflows

A year ago "using AI to code" meant autocomplete. Today my actual job looks more like being a founder with a small engineering team — except the team is agents, and I'm the one setting scope, reviewing diffs, and deciding what ships.

What changed

The unlock wasn't a smarter model. It was treating agent work like real delegation:

• Write the spec before the code. An agent given a clear spec produces something I can review in minutes. An agent given a vague prompt produces something I have to rewrite.
• Batch the work. I run multiple independent projects in parallel — a property analytics SaaS, a client site build, a trading bot backtest — because each one is scoped tightly enough that an agent can own a full slice without me babysitting it.
• Review like a founder, not a linter. I'm not checking indentation. I'm checking: does this match what the business actually needs, did the agent cut a corner that will bite later, is this the simplest version that works.

What it's good for (and what it isn't)

Agentic workflows are excellent at turning a well-scoped idea into a working build fast — a landing page, a backend service, a migration. They are not a substitute for product judgment. The scoping conversation — what to build, for whom, and why — is still entirely mine, and it's the highest-leverage part of the job now that implementation is cheap.

If you're building solo and haven't restructured your workflow around this yet: start by writing specs like you're handing them to a junior engineer you trust but don't want to micromanage. That single habit change does more than any tool choice.