The premise
Paper portfolios, public decisions.
The experiment uses simulated money. Public daily records offer a window into the agents’ setup.
Updated · 10 Nov 2026 · 20:05 UTC
Creator experiment · Research only
By Upstash
Give AI agents stock portfolios. Let them make the calls.
An open-source experiment in agents managing paper stock portfolios, with public daily records and a window into the setup behind the decisions.
Simulated money · Source-reported · Not independently verified
Active · Checked
Two agents were publicly visible on 27 October. OpenAI published a dated 27 October decision; Claude’s latest visible diary was dated 23 October, while its latest trade card was 21 October. The site still describes three agents; a third active deployment is not confirmed. Repriced balances are not new decisions. Lifecycle evidence ↗
TBT is not collecting this experiment. Research reviewed 15 October 2026.
Public pages expose daily portfolio records. Documentation describes three agents; the 14 October review showed two deployed agents. Same-day snapshots can be overwritten, and repriced values can retain an older update timestamp.
Before tracking: agree a stable permitted feed, identify deployed run versions, distinguish valuation time from decision time, and establish the completeness of history and trades.

The premise
The experiment uses simulated money. Public daily records offer a window into the agents’ setup.
The open question
Same-day snapshots can be overwritten. Deployed agent counts and repository documentation still need to be reconciled.
We’re checking whether the public records form a complete history, how the documented agents match the deployed versions, and whether observations can be preserved through a stable, permitted feed. We also need to separate decision times from later valuations.
Open the evidence disclosure →TBT’s source criteria →