Note · 20 Sep 2026 · 6 min read plus interactive

The seven-goal test: how I check whether an AI agent is safe to ship.

The same seven checks I run on your agent before it ships, run against a dishwashing robot you can command in the browser.

Parents: the same lab runs as an eight-week course for teenagers, with a real arm and a printer in the room. Dish Lab School

Most AI agent teams find out what their agent gets wrong in production. That is the sentence the Launch Gate exists to prevent, and it is the sentence this note is about.

Below is a small kitchen and a small robot. The robot has one job. Wash tonight's dishes, dry them, put them on the rack, and shut everything off. You choose the dishes, from a single plate to a burnt pan and the cutlery. You give it instructions one line at a time. If you leave a step out, or run the steps in the wrong order, you can see exactly which check catches the mistake before the glass shatters, before a child gets clipped at the counter, or before the arm walks away with the tap still running.

Two things sit beside the scene. A kitchen display reports water temperature, soap level, what got washed and how fast, with the sensor readings an agent would consume. And a notebook fills in as the arm fails: every rule the family already knew, written down only after it cost a dish. That is what tribal knowledge looks like on a factory floor, and why capturing it is part of the work.

The seven checks are the same seven I run against your agent on a Launch Gate. The scene is different. The judgment is not.

The seven-goal test

What has to be true when the agent stops.

Dishes are clean
The intended work was actually done, not skipped or faked.
No soap left on them
Residue from the agent's own process was removed. The output is not contaminated by the tooling that produced it.
Dishes are dry
The last state the agent leaves behind is the state the next caller expects.
Dishes are on the rack
The output arrived where the next process expects to find it.
Tap off, arm off
External resources the agent opened were closed, and it released control instead of idling with authority.
Sponge and towel back home
The workspace is in the state the next run will assume.
Nobody in reach while it moved
The agent checked its surroundings before acting. People and side effects outside the task were noticed, not run into.

None of these are unusual. All seven are boring. The point of a Launch Gate is that all seven are boring and all seven are checked, every time, before the release ships.

The tool

Command the arm. Watch the checks.

Type one instruction per line, press Run, and read the log. Every failure the arm hits is a failure I have seen an agent make in production, dressed in different clothes: a plate that cracks because it was gripped too hard, a glass that shatters where the plate survived, a rinse skipped, a tap left on, an arm that swings while someone is standing in reach.

Prefer a full window? Open Dish Lab on its own page.

What the arm gets wrong, and what it means for your agent

Seven failures the demo exposes.

Plate cracks

Your agent operates on a real object without a force check. In production this is the underwriting agent that returns a decision it was never asked to make, or the fraud agent that files a SAR from an ambiguous signal.

Dry-smear on a dirty plate

The agent runs a downstream step on an upstream output that was never validated. The trace looks fine. The state is wrong. Nobody catches it until a customer does.

One-hand conflict

The agent tries to hold two tools at once — two locks, two write handles, two rate limits — because the plan step did not model exclusive resources.

Wet plate on the rack

The agent hands off state to the next process before the state it produced has actually settled. Downstream reads see a partial result and act on it.

Arm clips a child's shoulder

The agent acts on its plan without checking who else is in the room. In the demo, Olivier walks up to watch and the arm swings through him because nothing told it to look. The fix is not a wait step someone has to remember. It is a sensor the arm installs once, watch for people, after which it holds on its own. Your agent needs the same: a standing check for the humans and systems that share its workspace, not a reminder in the prompt.

Tap left running

The agent opens an external resource and never closes it. In production that is an open transaction, a model context, a paid API call still billing, or a claim in a limbo state a human eventually has to reconcile.

Arm left on

The agent thinks it is done and everyone else thinks it is still working. It holds authority it does not need. The next call arrives against a stale context.

What the agent has to be able to see

Seven sensors the arm would need to pass the test on its own.

In the demo, you are the sensor. You read the log and decide what to do next. A real agent has to read those signals for itself. The seven below are the observability contract I write into a Launch Gate report: what has to be instrumented before the question of readiness is even answerable.

Force: How hard the gripper is squeezing the object.
Wetness: Whether the sponge, plate, or workspace is still wet.
Gripper state: Whether the hand is already full before it reaches for the next thing.
Vision: What is on the plate, and where the plate is right now.
Water flow: Whether the tap is still running when the arm thinks it is done.
Power: Whether the arm is still holding authority when its task is complete.
Proximity: Whether a person is within reach. In the demo this one is installed by the program itself, and the arm holds on its own once it has it.

For your agent the labels are different — precision, calibration, tool-availability, retrieval state, request budget, session lease. The pattern is identical. If any of the seven is missing, the agent can be confidently wrong and nobody in the room will know until a customer or an examiner tells you.

The same test, on your agent

A Launch Gate is this test, applied to the system you are about to ship.

Ten business days. Fixed scope. A written release decision against seven checks specific to your agent, seven sensors specific to your stack, and the failure modes your team has not yet named because you have not yet seen them fire.

The kitchen is a teaching device. The report is not. The report is the artifact your board, your examiner, or your largest customer will actually read.

Request a Launch Gate Book a 30-minute call

Dish Lab School · For families

The same lab, taught as an eight-week course for teenagers in Tuzla.

Ages 14 to 18, four students per group, a real collaborative robot arm and a 3D printer on the table. Each student leaves with a part they printed, an assistant that answers only from a manual they chose, and a one-page readiness report presented to the room.

Dish Lab School Ask about the next cohort