Choose a harness

For: plugin authors.

Use a harness plugin to run an existing agent runtime as one process per job on a node. Each job becomes one session of that harness.

For example, the Claude Agent SDK supplies file reading and editing, search, command execution, skills and subagents. A harness adapter makes those capabilities available within Orchestrator Zero's job controls.

Use the Claude Agent SDK

The official harness plugin is harness-claude. Install it on the nodes that should run harness agents, and set the tenant's Anthropic key once:

oz0 plugin install <git url of the harness-claude plugin>
printf %s "$ANTHROPIC_API_KEY" | oz0 secret set ANTHROPIC_API_KEY

Then give an agent, in any plugin, runtime: harness:claude-agent-sdk:

name: acme-coder
description: Makes small changes to code and runs the tests.
runtime: harness:claude-agent-sdk
model:
  use: smart                          # a tenant alias; any provider record that takes Anthropic's Messages API
instructions: prompts/acme-coder.md   # added to the SDK's own system prompt
limits:
  max_steps: 30                       # turns
  timeout: 30m
harness:                              # options for the harness, passed on as they are
  tools: [Read, Write, Edit, Glob, Grep, Bash]   # the default; anything else is refused
  permission_mode: acceptEdits        # default, acceptEdits, plan or bypassPermissions
  system_prompt: append               # or replace, to use only the instructions
  max_budget_usd: 2.0                 # the SDK stops the session at this cost

Run it like any agent: oz0 run acme-coder "Fix the failing test in tests/test_dates.py". Its text and tool calls appear in the live stream, and its model calls are metered under the job.

What a session gets

An empty working directory

The session gets its own working directory, removed when it ends, and a private state directory for the harness's configuration. A session without a sandbox (tier 0) also gets a temporary folder of its own (TMPDIR), removed with them, so what its programs leave there does not stay on the node. It is a short folder in the node's temporary folder, not under the session's, so that a Unix socket a program binds there stays within the limit on a socket's path. A node killed in the middle of a session leaves these folders behind (temporary session files).

• No user, project or local settings, no memory between sessions, no telemetry and no auto-update.
• The node's LLM proxy as its API, so the tenant's key stays in the edge, with the job and agent on every call so the gateway meters the session.
• A minimal environment: never the Temporal proxy's token or a provider key.
• Its agent's skills, as a read-only folder beside its own folders, outside the working directory: the Claude Agent SDK loads it as a plugin of that session only, with the Skill tool (allowed for an agent with skills, and decided by the job like any call); an ACP harness's tool reads it where the harness's acp.skills says. A harness of your own gets the folder's path as skills_dir and the names as skills in its session request.

Related tasks

• Harness tool controls: Apply hooks, policy and approval to a harness session.
• Recover a harness session: Understand node loss and restart-safe sessions.
• Write a harness: Wrap a framework or implement the harness protocol.
• Use an ACP agent: Configure ACP harnesses and enforce their session controls.