Concepts
The concept pages connect the blog post, videos, experiments, and OpenHands implementation path.
Core Thesis
Agent=Model+Harness
Model capability and execution reliability are different things. A stronger model can still fail if the harness gives it vague context, wrong tools, no memory, unsafe permissions, weak verification, or no way to recover when it gets stuck.
The harness is the system around the model:
- instructions
- tools
- memory
- safety
- runtime architecture
- routing
- verification
- observability
Start With The Framing
- Harness engineering repo and presentation material
- Engineering the Harness video
- Why OpenHands
- Visual Intro
- Harness Levers
- Experiments
Then Build It
Move from concepts to the runnable project path. Each project changes one harness lever and asks whether the trace got better.