Home / Context and cost

A sub-agent costs before it has done any work

A subagent costs before it has done any work, because its conversation already contains items you did not write, loaded again at every launch. With no tools field it inherits every tool, and with no model named it runs on the same model as you. List its tools and name its model by role.

A subagent is a separate conversation that Claude Code opens for a task. Before its first instruction, that conversation already contains items you did not write, and every launch loads them again. A subagent's cost is therefore made of this startup and of the work requested.

What a subagent contains at the start

The subagents documentation lists, among other things, what it receives: its own system prompt with environment details, every level of CLAUDE.md files the main conversation loads, the full content of the skills named in its skills field, the MCP servers in its mcpServers field, and the start of its memory file if memory is on. For tools, the rule is the reverse of what one assumes: with the tools field omitted, the subagent inherits every tool available. The more items at the start, the heavier the first call.

A custom agent with listed tools

A file placed in .claude/agents/ defines an agent that Claude can launch by name. This one keeps only what its task requires:

---
name: reader
description: Reads files and reports what they contain, without changing anything.
tools: Read, Grep, Glob
model: haiku
---
You read the requested files and report the facts found, with the file name.
You do not judge and you change nothing.

The three listed tools are the only ones this agent can use, and model: haiku gives it a small model. The documentation specifies that Claude Code watches these folders and picks up an added or edited file within a few seconds; after creating the very first file in an agents folder, a session restart is needed.

Naming the model by role

With no model named, the documentation gives this order: the call's parameter, then the model field of the definition, then the CLAUDE_CODE_SUBAGENT_MODEL variable, then the main conversation's model. A subagent that received nothing else therefore runs on the same model as you, with nothing flagging it at launch; the /tasks command shows the model on the subagent's row. A mechanical collection (read, list, extract) can go to a small model; writing and analysis to a mid-size model; the judge, who must refute, to the most capable. The effort level is set the same way: the Agent tool accepts an effort parameter for a given launch, and an agent's definition can set its own.

Measure before concluding

The real weight of each item is read on your own setup, not in an average. The lesson on the bill shows where to read consumption by item; apply it to a default subagent, then to the same work given to the restricted agent.