Hey Product Hunt 👋 I’m building infrastructure for recursive self-improvement (RSI). My current belief is that getting there requires more than improving the model itself. We first need to expand the boundary of the agent harness. Today, most agents are constrained by the harness they run in: what tools they can use, what environments they can enter, and what kinds of tasks they can actually complete. That limits something equally important: the data they can learn from. If agents can operate across more real-world work environments, they can encounter a much wider distribution of tasks, failures, feedback, and trajectories. That creates the data needed to improve not just the agent, but eventually the underlying model. That’s why we built Hitch. Hitch is an open-source infrastructure layer that lets agents connect to and choose between different harnesses. Instead of treating the harness as a fixed container around the model, we want to make it something the agent can select and eventually optimize. The longer-term loop we’re interested in looks like this: broader harnesses → more real-world tasks → more diverse trajectories → better models → better harnesses Ultimately, I think models and harnesses need to co-evolve. Hitch is a small first step toward that. We’d love to hear from developers and agent researchers: what environments or harnesses would you want an agent to be able to enter next?

Hitch

Let agents choose the right harness for every task

Let agents choose the right harness for every task