The Claude Code leak made one thing obvious:
The model is only a tiny part of an AI agent.
Most of the system is harness: the agent loop, context, tools, permissions, sandboxing, evals, and memory.
I wrote a survey mapping this field: harness engineering.
(link in reply)
The Claude Code leak made one thing obvious:
The model is only a tiny part of an AI agent.
Most of the system is harness: the agent loop, context, tools, permissions, sandboxing, evals, and memory.
I wrote a survey mapping this field: harness engineering.
(link in reply)