Research lab
Studying agents that do real jobs.
Most AI research tests agents on puzzles. We study them on payroll weeks, upset customers and twenty-year-old codebases, and publish what we learn.
Client data is never used for research without written consent.
Four open questions
Knowing when to stop
Can an agent tell when it’s out of its depth?
Calibrated confidence, and handing over to a person at the right moment.
Exams from history
How do you test an agent on a business that’s never been tested?
Turning past tickets, bills and incidents into fair, leak-free evaluations.
Rules as code
Can a policy manual become a guardrail?
Compiling plain-English business rules into checks that hold up, even against prompt injection.
Crews, not soloists
When do six agents beat one?
Handoffs, shared memory and the failures that only appear when agents work together.
Field notes
Anonymised write-ups from real installs. The first five are in progress.
Accounts payable
The bank-change letter that almost got paid
Why invoice fraud detection works best against each supplier’s own history.
Evaluation
Two weeks of shadow mode
The gap between exam scores and live performance, and how to set a go-live bar people trust.
Voice
A phone agent that hands over gracefully
Triage scripts, warning words and why the first eight seconds matter most.
Security
Injection in the inbox
Hidden instructions found in real supplier emails and PDFs, and the defences that held.
Legacy code
Parity tests first, port second
Recording what production really does before touching a line.
Work with the lab
We partner with universities and industry on applied agent safety, including PhD placements and joint projects. Open-source tools for agent exams, policy guardrails and shadow-mode comparison are planned.
hello@teammate.autos