Ponder Fix-it
Your agent fixed it. Put that experience to work.
Continual learning starts with real mistakes and real corrections. Connect your favorite harness, collect useful recoveries, and train a specialist that can help with the next tool call.
From a failed tool call to a better next attempt.
- 01
Connect your harness
Copy the setup prompt into your coding agent. Connect your OpenPond account, select the model and project, and review what the integration can capture. Start with OpenPond or Claude Code.
- 02
Capture what actually worked
When a tool fails and a later attempt fixes it, keep the original call, relevant context, correction, and observed result together. Review the recovery before it becomes training data.
- 03
Train, compare, then choose
Use approved examples for supervised fine-tuning. Compare the candidate against the base model on held-out failures and already-correct calls. Activate a version only after it earns its place.
- 04
Bring help back to the next session
Your agent can consult the specialist for advice about a tool call. The agent still runs its own tools with its normal permissions and verifies the outcome. A consultation never bypasses your approval flow.
Keep your main model. Keep your tools.
Fix-it adds a focused consultation tool to your workflow. It does not require moving your project into a new sandbox or replacing the agent you already use.
You choose what becomes training data.
Captured evidence belongs to your project. Review useful examples, revoke an installation, and remove retained events through your workspace controls.
Evaluation decides what goes live.
Training can produce a worse model. A finished run is not an automatic promotion: compare recoveries and valid-call controls before selecting a learned version. Existing-model consultation is available while you qualify it.
Pay for the work you use.
Use your OpenPond balance for hosted usage. Consultation and training have separate meters, so you can try the integration before approving a training run.
- Open-source integration
- Free to install
- The harness integration has no install fee. Your main agent’s model usage remains on its usual account.
- Specialist consultations
- Per token
- Input, cached input, and output use the selected model’s current rate. Check that model’s rate before use; there is no single fixed Fix-it rate per million tokens.
- Training and serving
- Usage based
- Approve a bounded training run before compute starts. GPU training and hosted serving are billed separately from consultation tokens at their applicable rates.
Start with your next useful correction.
Connect your harness and see the evidence before deciding what to train.
Get API Key