Docs - Context
Context and token estimation
Every agent turn stacks six layers of context. See exactly what fits in each one and how many tokens it costs.
1. The six layers
Agent context is built in order. Each layer is measured separately.
| Layer | What it holds |
|---|---|
| System | The agent's core identity and behavior instructions — set once when the agent is created. |
| Tools | Tool definitions — the names and descriptions of every tool the agent can call. |
| Instructions | Per-turn instructions you've written into the system prompt — updated before each turn. |
| Knowledge | All files you attached to the agent from your knowledge library — injected into the prompt. |
| History | All prior messages in this conversation — grows as the user and agent exchange turns. |
| Message | The user's current input — the thing the agent is about to respond to. |
2. Token estimates
Every layer carries a token estimate: the number of tokens the layer will consume in the model's input. We estimate tokens as ceil(characters ÷ 4) — the same heuristic the runtime uses for budgets. Larger files, longer history, or more tools push the estimate up. You can read the exact estimate for each layer on every turn.
3. Reading a snapshot
After each turn, a snapshot of the context layers is available in the inspector. You can see the character count and token estimate for every layer, so you know exactly what went into the model that turn. This is useful for debugging why an agent isn't seeing certain information, or why a conversation is suddenly costing more tokens.