# Output-first agents

The best way I've found to build an agent, or to decide not to.

## The claim

An agent's value is doing work on a human's behalf. Two forms: suggesting things a human
approves, or eliminating a workflow entirely. Everything else (chat interfaces, dashboards
with a sparkle icon) is decoration.

If that's true, then the riskiest assumption in any agent project isn't technical. It's
whether the work the agent does is work anyone wants done. And you can test that
assumption without building the agent.

## The method

1. **Get data as real as you can.** Production data, anonymized if needed. Synthetic data
   validates nothing because customers can smell it.
2. **Generate the output first.** Whatever the agent would produce for this customer (a
   collections plan, a prospect list with quotes, a filled-out form), make that artifact
   now, by hand plus model, before any application exists. A PDF is a fine container.
3. **Pitch the output, not the product.** Put it in front of a real customer: "here's
   what you'd get if you gave this thing your data." They will tell you three things fast:
   is it correct, is it valuable, and what would they never let a machine do unattended.
4. **Iterate at output speed.** Regenerating an artifact takes hours. Rebuilding an app
   takes months. Do your disagreeing at the cheap stage.
5. **Build the machine last.** By the time you write application code, the demand is
   validated and the human-approval boundaries are mapped.

## Where it came from

Partly from shipping ML before the hype and grading my own work honestly. An early
recommendations model taught me the gap between "the model works" and "the model makes
money." A 2024 order-prediction feature taught me that interesting-to-clients is not
valuable-to-users, and that lesson cost a quarter of engineering time I'd rather have
spent on a validated bet. The collections agent, run output-first, reached its first
verbal commitments before its web app existed.

The pattern generalizes past agents: it's just demand validation with the cost of the
prototype driven to almost nothing. But agents are where the economics get extreme,
because the output IS the product. Nobody buys a collections agent. They buy collected
receivables.

Alvaro Soltero · [ajsoltero.com](/)