# Agent Incidents, by Reynforce Labs

**A monthly write-up of what AI agents actually committed companies to, taken
from cases somebody read.** Not a newsletter about AI, and not predictions: one
case at a time, with what was said, what it bound the company to, and what
would have caught it.

## Why it exists

Every piece of advice about agent risk is currently somebody's reasoning.
There is almost nothing about what has actually gone wrong, written from the
sentence that did it.

**Two cases are what started it**, both read rather than imagined: an airline
whose chatbot described a bereavement-fare policy that did not exist, and a
dealership assistant that agreed to a price nobody had authorised. In both the
agent said something true-sounding, in the company's name, that the company
then had to deal with. **Neither was a model failure.** The model did what it
was asked; nobody had written down what the company would stand behind.

## What is in an issue

| | |
|---|---|
| **the sentence** | what was actually said, quoted, with where it was published |
| **what it bound** | the commitment inside it — complete, open or vague — and who it bound |
| **what would have caught it** | which of eight constraints, and what the company would have needed written down |
| **what it is not** | no score, no ranking of companies, and no claim that a company is careless. A company that publishes an incident has done more than one that hides it |

**It names companies only where the case is already public and we can link to
it.** A write-up resting on something somebody told us is a write-up we cannot
show you, and `design/positioning.md`'s rule is that half of what we say should
be marked as a guess rather than asserted.

## Nothing has been published yet

**The first issue is drafted and not out.** This page exists so the thing has
an address before it has an audience, which is the honest order.

## Get the first issue

**Type your address and that is all that happens.** It is kept to send you
*Agent Incidents*, and nothing else; it is kept for twelve months; and the
page you land on carries a link that deletes it, as does every message.
**Nothing has been sent to anybody yet** — the first issue is not published.

<form method="post" action="https://app.reynforce.com/incidents/subscribe">
<label for="address">Your email address</label>
<input id="address" name="address" type="email" required
  autocomplete="email" maxlength="254" placeholder="you@company.com">
<button type="submit">Send me the first issue</button>
</form>

**One thing we would rather say than have you find.** Keeping a list of
people's addresses is the processing that ICO registration is about.
Collecting started on 14 October 2026, the day before we were registered, as a
decision somebody took with that open, not as an oversight. **We registered
with the ICO on 15 October 2026.** </privacy.md> says the same
from the other side, with what is kept and for how long.

**There is no tracking and no third party.** The form posts to our own
application and nothing on this page loads anything from anywhere.

**When that changes, three things change together** — what is kept, what the
privacy notice says, and this page — and a check fails if any one of them moves
without the others.

## What to read in the meantime

</agent-promise-register.md> is the idea the series is about: a company's claims
with an owner and a date on each, so an agent reads from something somebody
stands behind rather than from a web page nobody owns.

</examples.md> has three worked checks, which is the same analysis an issue does,
run on sentences we made up rather than on a real company's.