---
title: 'Solving almost every problem with AI: the honest version'
excerpt: 'The friction from idea to working software collapsed. The friction from working software to a system you can trust in production did not.'
date: '2026-06-28'
lastModified: '2026-07-04'
author: 'Teo Deleanu'
authorAvatar: '/team/teo.jpg'
tags: ['Agentic AI', 'AI Strategy', 'Production AI', 'AI Engineering']
keywords:
  - 'agentic AI for business'
  - 'production AI systems'
  - 'AI agents for small business'
  - 'AI prototype vs production'
featured: true
tldr: 'The cost of getting from an idea to a working prototype fell by roughly an order of magnitude. The cost of making that prototype correct, safe, and maintained in production barely moved. The winning method is not "point an autonomous agent at your life" but "find a narrow, expensive, measurable loop and close it with the smallest agent that does the job, an eval in front and a human behind."'
keyTakeaways:
  - 'What collapsed is idea-to-prototype cost, not idea-to-production cost. Those are different problems.'
  - 'Unattended, fully autonomous agents running a business A-to-Z is the part of the pitch that does not survive real customers.'
  - '"Free because open source" ignores the largest recurring line item: compute, API calls, and integrations all meter per use.'
  - 'The repeatable play is a narrow, expensive, closeable loop, with an eval gate in front and an accountable human behind.'
---

There is a video going around right now that says you can buy a Mac Mini, download a free open-source agent, tell it "start me a business," and wake up to a company running itself while you sleep. No money, no team, no technical background. A great reset. Move this week or miss it.

The core observation under the hype is real. The specifics are mostly wrong. Here is the part that is true, the part that is being sold to you, and the method that actually works.

## What actually changed

The cost of going from an idea to a working prototype dropped by roughly an order of magnitude. That is the whole shift. A single person running agentic coding tools can now produce, in a weekend, what used to take a small team a quarter.

![Two cost curves from 2020 to 2026: idea-to-prototype collapses around 2023-2025 while prototype-to-production stays flat. The gap between them is the work, and the business.](/blog/cost-curves.svg)

I run this. The doc-intelligence demo on this site is a real pipeline: you drop in a document, the system parses and indexes it, and a live model answers questions about it. You can try it in your browser. One engineer built and shipped it by running coding agents in parallel.

The rule at the end of this article is "if you cannot measure it, you cannot claim it," and it applies to me first. The demo's receipts:

| Receipt | The number |
| --- | --- |
| Built by | One engineer running coding agents in parallel |
| Timeline | Design spec May 29, 2026; live on Fly.io June 27, 2026; 74 commits |
| Infrastructure | Three shared-CPU Fly machines: API (1 vCPU, 512 MB), worker (2 vCPU, 2 GB), Postgres (1 vCPU, 1 GB) |
| Infrastructure cost | About $20 a month at Fly.io's published rates, running 24/7 |
| The meter | Gemini API at published rates: $0.15 per million tokens embedded; answers at $0.30 in, $2.50 out per million |
| Answer latency | About 9 seconds end to end for a grounded answer, measured during development (the model spends most of that thinking) |

Notice what the receipts say and what they do not. They prove the cheap half of this article: a real pipeline for the price of a coffee subscription. They do not prove the expensive half. This demo is exactly the kind of young system I tell you not to trust with a business yet; the difference is I am on the pager for it, and nothing rides on it but my credibility.

So the idea-to-prototype collapse is not marketing. It happened.

## What did not change

The cost of going from a working prototype to a system you can trust in production did not collapse. It barely moved.

Here is why, mechanically. An agentic system is agents calling tools calling services, with the model as the unreliable part in the middle. The model drifts between versions. It hallucinates tool calls. It fails silently and confidently. One LLM call that hangs for 90 seconds stalls every request behind it, and the queue backs up faster than autoscaling can react. None of that shows up in a 48-hour demo. All of it shows up the first week real customers touch it. I published [the demo's own week-one incident log](/blog/it-runs-itself-first-week) as the receipt.

The viral pitch skips this entirely. "Runs a business A to Z, unattended, with its own credit card" is the exact thing nobody who ships production agents actually does. You put an eval gate in front of the model: automated tests that catch it getting worse before a customer does. You put a timeout on every call. You keep a human on the hook for the consequences. The demo that runs for two days in a screenshot is not the system that runs for two years under load.

And "free because it's open source" ignores the meter. The weights or the API calls, the compute, the SMS and email and phone integrations all cost money per use. The recurring bill is the largest line item, and it is the one the pitch never mentions.

## The one real example in the whole pitch

Buried in the hype is a genuinely good play, and it is worth keeping.

Go to a business that leaks money in a narrow, boring place. The example given is an HVAC company that misses after-hours calls, so the caller reaches a competitor first. In a few days you can stand up an agent that answers those calls, sends an immediate text, produces a quote in near-real-time, and writes the result into their existing CRM. You measure the recovered calls. You raise their revenue by a measurable percentage. You become the person they pay to fix the next problem.

This is the one part of the pitch that survives contact with real customers. Notice what makes it work, though, because it is the opposite of "solve every problem":

1. The problem is narrow. One leak, not the whole business.
2. The problem is expensive. Missed jobs are real dollars, so the fix pays for itself.
3. The loop is closeable. Call comes in, text goes out, quote is sent, CRM is updated. You can measure whether it worked.
4. A human still owns it. You watch it, you fix it when it breaks, you are accountable for the output.

That is the method: find a bounded, expensive, measurable loop and close it with the smallest agent that does the job, with a check in front of it and a person behind it. "Point an autonomous agent at your life" is nowhere on the list.

<!-- CTA -->

## The honest method

Every problem that AI can actually solve today fits the same shape. The playbook:

- **Find a narrow loop that costs real money** *(your job)*. Not "transform the company." One process, one leak, one number that moves.
- **Build the smallest agent that closes it** *(engineering)*. Resist the urge to make it general. General agents are how you get silent failure at scale.
- **Put an eval or a guardrail in front of the model** *(engineering)*. You need to detect drift before the customer does. A regression that flips an eval score should stop a deploy before it reaches production.
- **Keep a human accountable** *(your job)*. Someone owns the output when it breaks, because it will break.
- **Ship, measure, iterate** *(both)*. The number you moved is the proof. If you cannot measure it, you cannot claim it.

Two of those five are engineering. The other three are judgment about your own business, and you are already the expert in that. The engineering you can hire or learn.

Do that ten times across ten businesses and the revenue the video promises starts to be reachable. The "unattended" part never arrives: ten loops means ten systems someone watches, patches, and answers for. You get there by closing loops you can measure, not by trusting a black box with a credit card.

## Where this leaves you

"Almost every problem is solvable with AI" is closer to true than it has ever been, for one specific reason: the cost of implementing a solution dropped far enough that problems which were never worth solving are now worth solving. That is real, and it is a big deal.

It also means you are not late. The video says move this week or miss it. The window here is measured in years, and it rewards the people who close measurable loops over the people who rush. The urgency is the sales tactic, not the situation.

The catch the pitch hides is that "solvable" still means the same engineering it always did, just cheaper and faster at the front end. Someone still has to put timeouts on the calls and gate the evals, and someone has to own the failures. The friction moved from "can we build it" to "can we make it correct and keep it correct." That second part is the work. It is also where the money is.

We do both sides of this. We build these systems for companies that need the loop closed now, and we teach engineers to build them for themselves. The demo on this site is one closed loop you can try in your browser. Start there, then bring us the loop you want to close.