Agent Context is the information made available to the model at a specific point during an agent's execution.
It can include:
- User input
- Conversation history
- Relevant memory
- Retrieved knowledge
- Tool results
- Current state
- System instructions
Context is not necessarily the same as state or memory.
State describes what the agent is doing; context describes what the model can see.
For example, an agent may have thousands of records in its memory, but only retrieve a few relevant ones into the current context.
Good context is the smallest amount of relevant information that enables the agent to make the right decision.