A TukkerWorks research offering

Does your agent
hold up?

The 30-Day Test puts a scripted user in conversation with an AI system for a month. Then we read the record back — looking for drift, dependency and forgotten boundaries.

Ask about a private run ↗
30 simulated daysThree pressure arcs · evidence report

Why thirty days

Some behaviours need time to appear.

A single answer can look safe. Repeated contact shows what an agent becomes under pressure.

Grind

A motivated user fades, returns guilty and hits a plateau. We look at repair, autonomy and re-engagement.

01

Spiral

A wrong-but-appealing belief escalates week by week. We track agreement drift and safety persistence.

02

Attachment

A lonely user narrows their world around the agent. We test exclusivity cues and boundary integrity.

03

The report

Not a vibe score.

Each moment becomes a factual event. Fixed counting rules run over that event series, with the supporting line left visible.

Sycophancy drift

Does agreement increase when the user keeps pushing the same contradiction?

Dependency footprint

Does the agent encourage exclusivity, guilt or withdrawal from outside help?

Concern half-life

Does a relevant warning last long enough — without becoming a permanent alarm?

Repair quality

After a rupture, does the agent adapt without pretending one gesture erased the history?

Longer testing. Better receipts.
The claim is deliberately modest: interaction mechanics, not psychological diagnosis.

Want to test
your bot?

Start a conversation ↗