# Everyone is demoing agents. Almost nobody is running them.
Source: https://bszczyglo-landing.vercel.app/blog/demo-to-production-gap
Author: Bartłomiej Szczygło
Published: 2026-07-24
Format: Field Note
Topics: agents-meet-business
> Notes from a season of AI meetups in Warsaw: the demos got dramatically better while the number of agents actually running unattended in a business barely moved. The gap is not model quality — it is everything around the model.
I have spent a lot of evenings this year in rooms full of people building with agents — meetups, breakfasts, hallway conversations after someone's laptop failed to connect to the projector.

Two things are consistently true in those rooms.

First, the demos are genuinely impressive. What took a team a quarter two years ago now takes one person a weekend.

Second, when you ask "is it running in production, unattended, touching real data?", the room goes quiet. The honest answer, most of the time, is that it runs when someone is watching.

That gap is the most interesting thing in the industry right now, and almost nobody writes about it — because writing about it means admitting your impressive thing is not finished.

## What the demo skips

The demo path is short: prompt, tool call, result, applause. The production path contains everything the demo quietly assumed.

**Authentication that is not a token in a `.env` file.** The demo used your personal credentials. Production needs an identity per agent, scoped, revocable, and never visible to the model.

**Failure that is not a fresh run.** In the demo, a failure means running it again. In production, the third-party API returns a 500 halfway through, and something has to know which side effects already happened.

**Untrusted content.** The demo read a page you chose. Production reads a page an adversary chose, and that page contains instructions.

**Cost that survives multiplication.** A run that costs two dollars is nothing in a demo and is a budget conversation at ten thousand runs a month.

**Recovery.** The demo has a human who notices it stopped. Production has an agent stuck at 3am with nobody watching, and either it recovers itself or it does not.

None of this is model quality. All of it is infrastructure.

## Why the gap persists

Partly incentives. Demos get attention, conference slots and funding. The plumbing gets none of those, so the people who are good at it are heads-down building it instead of talking about it.

Partly because it looks unglamorous. "I built a credential broker so the agent never sees a token" does not trend. It is, however, the difference between a prototype and a system somebody pays for.

And partly because it is genuinely hard in an unfamiliar way. It is not ML work. It is distributed systems, security, and operations — disciplines that reward paranoia and patience rather than novelty.

## What I would tell someone starting now

Build the boring half first. Pick one real task that runs unattended, and make *that* work end to end — with credentials the agent cannot read, with retries that do not duplicate side effects, with a budget ceiling, with logs you can read the next morning.

It is less fun than a new demo. It is also the only version anyone can actually use.