Open-source harness
The whole learning loop, in one open-source repo.
Run agents beside real work, preserve the evidence, evaluate what happened, and improve the harness or train a model. Desktop, CLI, TUI, Tasksets, and the training workbench remain inspectable and extensible.
Own the runtime, the evidence, and the path to improvement.
OpenPond keeps agent source, work traces, Tasksets, graders, model recipes, and deployment decisions together. Run locally without an OpenPond account, extend the contracts in source, and add hosted execution only when the work needs it.
- 01
Run real work
Use Desktop, CLI, TUI, or Web with files, terminals, browsers, subagents, and explicit human approval.
- 02
Make results measurable
Turn selected outcomes into source-backed Tasksets with explicit splits, verifiers, graders, and baselines.
- 03
Improve what matters
Refine the agent and harness first, then train a task-specific model when evaluations show that training is worthwhile.
Evidence, not anecdotes
Every improvement has receipts.
Keep the task, execution environment, trace, reward, grader, model version, and final decision connected. Reproduce the run and inspect the exact source that produced it.
Fork it. Extend it. Keep it yours.
The harness is built as ordinary source, not a hidden workflow graph or a collection of saved prompts.