OpenPond Harness Is Finally Ready to Try (v0.0.32)

OpenPond Harness is finally stable enough that I can recommend trying it: v0.0.32.

It is still alpha. Expect bugs. But the core loop is real now.

The goal

Use the work already happening in your conversations to explore RL environments, run evals, and ship better Agents, Skills, harness extensions, and RFT models.

If you are suffering from “what do I build now?” syndrome, my answer is RL environments. They are still under-explored, new papers are dropping every day, and they get much more interesting when they are grounded in work you actually repeat.

Then invite your non-dev teammates and ship them agents instead of another dashboard.

Conversations become evals; evals improve an agent, skill, model, or the harness itself.

What works today

• Codex app-level UI/UX, BYOK, subscription-backed models, and subagents.
• Agent Lab works with the Make Agent flow: build an agent, eval it, improve it, and ship new versions.
• RL Lab currently works with Fireworks RFT.
• Team Chat for paid teams, plus an open, Discord-esque Community Chat.
• OpenPond Cloud for hosted coding and sandbox runs. Sandbox usage is paid, but you can still use your subscriptions while coding in the cloud.

Profiles start local, can stay local, and are normal Git-backed source files. When you are ready, the same harness can move to your team or OpenPond Cloud.

RL Lab training a task-specific model with Fireworks RFT.

We are also aiming for feature parity with the coding harnesses you already use. The difference is that OpenPond is built around continuously improving the harness itself—not just finishing the current chat.

Try it

You can also grab the desktop app from GitHub Releases.

The OpenPond Community lives inside the app.

To join the community, make an account on openpond.ai, then connect it under Settings → Account.

And give the repo a star: github.com/openpond/openpond.