Claude Code agents

A Claude Code agent runs on Claude Code instead of agt's own model loop. Like a Codex agent, it edits files and runs commands in a project directory, and it can use agt's messaging, memory and task tools.

Requirements

• Claude Code installed on the machine that runs agt (command -v claude).
• Claude Code signed in under the same user account that runs agt. agt does not store Claude credentials; it uses Claude Code's own login and settings.

agt runs Claude Code in headless mode (claude -p with stream-json), one process per turn, and resumes the agent's session on the next turn.

Create a harness

{
  "engine": "claude-code",
  "executable": "/home/me/.local/bin/claude",
  "working_directory": "/home/me/projects/tools",
  "sandbox": "workspace_write",
  "model": "opus",
  "reasoning_effort": "high",
  "timeout_seconds": 1800,
  "description": "Builds and reviews code"
}

Field	For Claude Code
sandbox	workspace_write: file edits in the project are accepted (--permission-mode acceptEdits), and other actions such as shell commands ask for approval. read_only: editing tools are disabled and every action asks.
model	A Claude Code model alias (fable, opus, sonnet, haiku) or a full model name. Empty uses Claude Code's default.
reasoning_effort	low, medium, high, xhigh or max, passed as --effort.

Web console: Resources → New resource → Harness, then choose Claude Code as the engine.

CLI:

agt harness resource create reviewer --file ./reviewer.json
agt harness resource test reviewer      # checks that the executable is Claude Code; runs no model

Then create the agent on it, as for Codex: agt agent create reviewer --harness reviewer --prompt reviewer.

How agt connects to Claude Code

• Instructions: the agent's role prompt and skills are added to Claude Code's system prompt (--append-system-prompt).
• agt tools: agt's messaging, memory, task and resource tools are offered to Claude Code as an MCP server named agt inside the agt process, so they appear as mcp__agt__…. Claude Code does not ask for approval for them; agt checks them itself.
• Approvals: when Claude Code wants to do something its permission mode does not allow, the request comes to agt's approval queue and follows the Harness approval mode; see Permissions. You can approve once, deny, or cancel the turn.
• Sessions: each agent keeps one Claude Code session per conversation and resumes it each turn. /new starts a fresh session.

What Claude Code agents can and cannot use

Available	Not available
Role prompt and skills	agt MCP resources (configure MCP in Claude Code instead)
resource_management and task_collaboration tools	local_execution (Claude Code has its own shell)
Messaging, memory and task tools	Response policy all (only mentioned and manual)
/stop, /new, /status, live preview of the reply	/compact, /rename, /skills, /mcp, /permissions

Things to know

• Files are shared. Agents using the same working directory edit the same files.
• Stopping is not undoing. Cancelling a turn stops Claude Code, but changes already made stay.
• Your Claude Code settings apply. Project and user settings, CLAUDE.md and hooks in the working directory are loaded as they would be in a terminal.