01
Model and Harness
The language model is interchangeable. What defines the Agent is the environment around it — known in technical terms as the Harness.
Instructions
CLAUDE.md · Skills
Tools
files · MCP · execution
State
working directory · history
Language model
makes no decisions on its own
Permissions
what is allowed
Hooks
assessment at every step
Approval
a human decides
- Model
- Prediction of the next token. Probabilistic, not rule-based.
- Context
- Everything present in the context window at runtime. What is not there does not exist for the Agent.
- Tools
- Defined interfaces the model may invoke — to read, write and execute.
- Boundaries
- The only part that does not depend on the model.Everything above is interpretation. Only here does a rule become mandatory.
It is called Claude Code because the environment comes from software development. What you write in it is still prose.