AI Agents & LLM Systems · 21 September 2026

Why most AI agents die on the way to production.

The demo works. The pilot works. Then it meets real people, real data and a Monday morning, and it stops working. The failure is almost never the model. It is everything around the model — and every one of these is a thing I have had to fix in something that was already live.

Che ShivaRoughly a six minute read

01 / The gap

A demo is a performance. Production is a promise.

Anything can be made to work once. One person, one machine, one carefully chosen question, and the person who built it standing close enough to restart it. That is a performance, and performances are easy.

Production is different in kind, not degree. Production means someone who did not build it, who does not care how it works, uses it at nine on a Monday while a customer waits — and it is right, and it is fast, and when it is not sure it says so.

Almost every failed AI project I have seen cleared the first bar and never understood there was a second one.

02 / What actually breaks

Four failures, none of them the model.

It is confidently out of date.

An agent will state a fact that was true three weeks ago in exactly the same tone it uses for something true this morning. There is no tremor in its voice. If nothing forces it to check, it will not check.

The running config is not the written config.

You change a setting. The file is correct. The thing that is actually running loaded its settings hours ago and has never heard of your change — so it fails quietly, falls back to an old default, and looks fine.

Nobody priced the running.

The build is the cheap part. What matters is cost per month at real volume, who maintains it, and what the bill looks like when it is used the way you hoped rather than the way you tested.

It has no idea what it is not allowed to say.

Left alone, an agent will quote a price you never set, promise something you do not sell, and repeat an internal number to a customer. It is not lying. Nobody ever told it where the edges were.

03 / The one nobody expects

The dangerous answer is the one that sounds certain.

A wrong answer that sounds unsure gets checked. A wrong answer delivered calmly gets repeated — to a customer, in a meeting, on a call you are not on.

This is the failure mode that costs real money, and it is not a model problem. It is a grounding problem. An agent that cannot point at where it got something will eventually fill the gap with something plausible, because plausible is what it is built to produce.

The fix is unglamorous. Point the agent at a source of truth it must read rather than recall. Make it say ask a person when the question goes past what it can show. And check what it believes against what is actually running, not against the documentation, because those two drift apart quietly and constantly.

04 / What to do instead

Start smaller than feels impressive.

One process. Measurable before and after. A named person who owns it when it is wrong.

01↗

Pick the boring process

High volumeLow stakes per eventObvious when wrongAlready measured

02↗

Ground it before you scale it

One source of truthMust read, not recallCite or deferRefuse rather than guess

03↗

Price the second year

Cost at real volumeWho maintains itWhat failure costsExit if it stops paying