Show HN: Harness Tuner – eval-driven tuning loop for agent tools, prompts, model
github.com
3 threads
I feel like there are always so many things I could try to improve my agentic systems but doing it manually never seems worth it to me. Can’t wait to try this out on my projects.
I'm not quite sold enough to invest in migrating to the Kiln ecosystem, but it looks to be a great concept and well done project.
makes sense. We're working on a 1 command "harness-tuner --source ./existing_agent" which should make it easier to try.
Oh, this actually re-writes tools? I assumed another prompt optimizer but looks like this is an actual harness tuner.
yeah exactly. The tool re-writes have been the biggest gains. A ton of agent issues are in the tool layer, so letting the loop optimize the tools makes instances of the agent much faster and more accurate.