Fresh-context agent

A fresh-context agent is a separate run of a coding agent that starts with none of the working session's context. It reads the work with read-only tools and returns its verdict in a fixed shape.

Finding it

• Subagents defined in the agent's configuration: Claude Code reads them from .claude/agents/.
• Headless runs from the command line: claude -p, codex exec.
• Review commands the agent ships with or the project adds, such as /code-review and /security-review.

Using it

• Hand it the work to read (a diff, a file, a spec) and a checklist, not the story of how the work was done.
• Read-only tools only: the tools list in a subagent's definition, or --allowedTools for claude -p.
• Output the loop can parse: a schema in the prompt, and --output-format json for claude -p.
• A limit on its run (--max-turns), so a stuck review ends.

Examples: Claude Code subagents, claude -p, codex exec; review commands such as /security-review, or /code-review from a plugin.