Five Context Components

With each LLM call, the context consists of five parts:

System prompt;
Tool definitions;
User messages;
Model responses (reasoning + content + tool_calls);
Tool execution results.

The first two are a static prefix, unchanged throughout the dialogue; the last three are a dynamic trajectory.

Ablation shows the indispensability of each:

without tool definitions the agent cannot act;
without tool results it loops;
without reasoning decisions contradict each other.

Related: Agent Cycle in Code, Chat Template, Context Engineering