rdi-berkeley/agents-last-exam ? reverse-engineered prompt

Build me a Python project for benchmarking AI agents on real work tasks.

I want a tool that can set up a sandbox, give an agent a task, let it work until it is done, then score the result with a hidden reference and save the logs. It should support running in a few environments, like cloud VMs, local containers, and a pre existing sandbox, and it should work with both terminal style agents and agents that can also click around a desktop.

Please include a small set of example tasks, a way to run an experiment from a YAML config, and a clear structure for adding new tasks and graders. I also want the run to collect a full trajectory, screenshots, tool calls, and output files so the run can be replayed later.

Make it feel like a real open source evaluation framework, with good docs, sensible defaults, and tests. If you need to, look up current docs online for the cloud and VM parts.