Skip to Main Content
Talk Intermediate MIT

Lessons from building an open-source agent harness

Proposal status is Rejected
Session Description

A field report from building orion-core, the open-source agent harness behind OrionPod. What running models locally taught me about prompt formatting, tool calling with small models, context pruning, structured events, and the bugs that actually bite, plus the one architectural seam I'd keep if I started over.

Key Takeaways

  • Different aspects of harness engineering.

  • Lessons from different approaches.

  • What works for bigger models, might not work for smaller models.

  • Build your own agent the right way.

References

https://github.com/anistark/orion-core

Session Categories

Introducing a FOSS project or a new version of a popular project
Talk License: MIT

Which track are you applying for?

Main track

Speakers

Kumar Anirudha
Kumar Anirudha Founding Engineer | Vibrant Labs

I build RL envs at Vibrant Labs and maintain an open source evals library called ragas.
I also built OrionPod as a side project and came along with it was orion-core, an open source agent harness. I also build dev tools like wasmrun, feluda, runek and so on.

https://x.com/kranirudha