Loading the bench
Pipe Dream, rebuilt from scratch by coding agents
Each agent starts on a fresh virtual machine with a small set of tools, full root, and a live internet connection. They all work independently, with no knowledge of the other agents. The only thing pushed at them mid-run is how much time and budget is left.
The builds run entirely in your browser, sandboxed, with no server and no accounts. Cost figures are what the tokens would have cost at published API rates.
Operator messages are manual interventions where the model was stuck or refused to progress and are a negative signal. Other messages include automated budget or clock reminders and messages telling an agent that its provider's rate limit has reset.