Connected agents
Preview
Connect agent activity and explicitly import completed conversation evidence.
Open Connected agents to configure where conversation evidence comes from. Supported source cards include Ponder Pal, Codex, Claude Code, OpenCode, Grok Build, Hermes, Pi, Oh My Pi, and OpenClaw. The OpenPond card switches between Web (Ponder Pal) and Local (OpenPond local). A custom Agent import is also available.
Connect a source
- Choose the source card and open its setup.
- Follow its source-specific collection or export instructions on the machine that owns the conversation.
- Choose the workspace and destination offered by the setup, and review the collection scope before enabling access.
- Inspect status and imported tasks before using them in an experiment.
For ongoing local collection, the source setup provides an openpond import connect command with your Project and workspace selected. Run it on the source
computer and approve the connection. A background collector uploads the chosen
history and keeps syncing new activity while that computer is online. Use
--once for a single import. The continual learning
guide has commands for every
supported local source.
Manual file imports use structured session exports. Preserve related files and explicitly selected branches or child transcripts when required. Finish the turn before using it as completed evidence.
Review before import
The import preview shows the selected conversation and missing context. Check that it contains the intended turn, tool evidence, and outcome. Exclude secrets and unrelated work. Importing records evidence; it does not approve that evidence for training or assert that the agent succeeded.
Use Ponder to inspect recorded activity and choose cases for a Dataset or Experiment.
Manage access and recovery
Open the source's details or setup controls for collection status and recovery. Use its connection controls to add a different computer or agent profile. Disconnecting revokes future access from that source; saved tasks remain. If an import reports incomplete context, correct the export or narrow the selection before treating it as reproducible evidence.