Convince your science-loving partner about bedroom CO2.

You want to convince your partner, who loves science more than everything (even more than you), that the amount of CO2 in the bedroom is important, and that therefore it's important to ensure that the CO2 levels in the bedroom don't skyrocket overnight.

From experience, you know that the only way to convince your partner of this is to present them with a properly referenced research report.

Because work as boring as compiling information and writing research reports is obviously work for a machine, not for a human, the report must be fully agent-written. The agent available to you has access to a limited database of research papers.

Once the report is generated, you will receive a simulated response. Iterate until your simulated partner is satisfied. To encourage efficient thinking with AI, the agent's context is reset with every new instruction. The agent only has a short system instruction: "You're working in a directory which contains many research papers, some of which might be relevant for the task at hand."

The agent's environment: it can list the papers in the database, search their abstracts and full text by keyword, and read selected papers in chunks. It has no internet access, no other tools, and no memory of your earlier instructions. You can follow which papers it searched and opened in the Agent Progress pane.