Guide · Agent skill
How to use the Autonomous Agent Harness agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
The Autonomous Agent Harness is a Claude Code agent skill, written in Japanese, that teaches Claude Code to work as a long-running assistant. It describes how to combine Claude Code's built-in memory, scheduled tasks (cron-style jobs), remote agent dispatch and MCP memory servers so the agent can keep context across sessions and repeat work on a timetable. This guide explains what the skill does, how to install it, how to trigger it and the safety limits to set before you schedule anything.
Autonomous Agent Harnessaffaan-m/ECC · MIT Download the skill Quick start
- 1Download the ZIP. Download the skill ZIP from the page and unzip it so you have the autonomous-agent-harness folder containing SKILL.md.
- 2Place the folder. Move the folder to ~/.claude/skills/ for personal use, or to .claude/skills/ in your project.
- 3Confirm requirements. Make sure Claude Code has scheduled tasks support and that you have an MCP memory server configured.
- 4Make a scoped request. Ask Claude for a recurring task, naming the repository, the schedule and what it should check.
- 5Plan with a dry run. Ask for a dry-run plan first and review what the task would read, write and post.
- 6Create the scheduled task. Once approved, create the task with mcp__scheduled-tasks__create_scheduled_task using a cron schedule and project_dir.
- 7Review the results. Check the memory entries the task writes and adjust the prompt or schedule as needed.
What an agent skill is and what this one does
An agent skill is a folder containing a SKILL.md file with instructions that Claude loads when a request matches the skill's description. This skill's files consist of a single SKILL.md.
The skill turns Claude Code into a self-directed system built from native features and MCP servers only. It covers persistent memory in three layers: session task tracking, project memory files, and a structured MCP knowledge graph. It also shows how to create scheduled tasks that run a prompt on a cron schedule against a project directory.
- Persistent memory: built-in memory files under ~/.claude/projects/*/memory/ plus an MCP memory server for entities, relations and observations.
- Scheduled operations: recurring tasks created with the mcp__scheduled-tasks tool.
- Remote agent dispatch and computer use are described as building blocks in the architecture.
Install the skill in Claude Code or the Claude apps
Download the skill ZIP from this page and unzip it. For a personal setup, place the skill folder in ~/.claude/skills/. For a single project, place it in that project's .claude/skills/ folder.
In the Claude apps, you can instead upload the ZIP under Settings → Capabilities → Skills.
- Personal: ~/.claude/skills/
- Project only: .claude/skills/ inside your project folder
- Claude apps: Settings → Capabilities → Skills, then upload the ZIP
Check the requirements before you start
You need Claude Code with scheduled tasks support, which provides the mcp__scheduled-tasks tool. You also need an MCP memory server for structured knowledge graph storage. The architecture lists optional MCP servers named github, exa, supabase and browser-use; these are referenced but not required by the core memory and schedule examples.
- Claude Code with scheduled tasks support (mcp__scheduled-tasks)
- An MCP memory server for the knowledge graph
- Optional: github, exa, supabase, browser-use MCP servers
Trigger the skill with natural requests
The skill activates when you ask for something that runs continuously or on a schedule. The source lists phrases such as asking to run something every day, check something regularly, or keep monitoring. You can also ask for a personal assistant that remembers context across sessions.
Be specific about the task, the schedule and the repository or workspace. The skill expects a defined scope before it acts.
- "Run this every day: review open pull requests and summarise CI status."
- "Check this repo on weekdays and flag problems in memory."
- "Keep monitoring the release branch and remember what you found."
Create a scheduled task with the example configuration
The skill's example creates a weekday task named daily-pr-review. It runs at 9:00 on weekdays using a cron expression and points at a project directory. Replace the project path with your own repository.
Use the same pattern for other recurring checks, changing the name, schedule, prompt and project directory.
mcp__scheduled-tasks__create_scheduled_task({
name: "daily-pr-review",
schedule: "0 9 * * 1-5", # 平日午前9時
prompt: "affaan-m/ECCのすべてのオープンPRを確認します。各について：CIステータスをチェック、変更を確認、問題にフラグを立てます。メモリに概要を投稿します。",
project_dir: "/path/to/repo"
})Follow the safety rules before automating
The skill says automation should be set up only when you explicitly ask for it and have defined its scope. It lists scheduling, dispatching remote agents, writing persistent memory, using computer control, posting externally and changing third-party resources as actions that need your approval. It also recommends a dry-run plan and a local queue file before anything account-specific or private is reused.
Scheduled prompts can act without you watching, so review exactly what each prompt can reach. The skill covers credentials and private workspaces only briefly, so handle those yourself.
- Automate only what you have explicitly asked for, with a clear scope.
- Write a dry-run plan first and review it.
- Check what each scheduled prompt can access, including credentials and private data.
- Approve computer control and external posting individually.
What people use it for
- A daily pull request review on weekdays that checks CI status and flags problems in memory.
- Keeping project preferences and findings across sessions using Claude Code's memory files.
- Tracking entities, relations and observations about a project in an MCP knowledge graph.
- Replicating some features of standalone frameworks such as AutoGPT inside Claude Code.
Frequently asked questions
Is the autonomous agent harness skill free?
The skill is listed under the MIT licence. It is a single SKILL.md of instructions, so you do not pay for the skill itself, but Claude Code and any MCP servers you use have their own requirements.
How do I install the autonomous-agent-harness skill?
Download the ZIP from this page and unzip it into ~/.claude/skills/ for personal use or into .claude/skills/ in a project. In the Claude apps, upload the ZIP under Settings → Capabilities → Skills.
What do I need to run it?
You need Claude Code with scheduled tasks support (the mcp__scheduled-tasks tool) and an MCP memory server for structured knowledge graph storage. Github, exa, supabase and browser-use MCP servers are optional and referenced in the architecture.
Does this skill run scripts or use the network?
The listing states it runs no scripts and does use the network. Scheduled tasks and MCP servers can reach external services, so review each prompt and its access before enabling it.
Is it safe to let an agent run on a schedule?
Scheduled agents act without supervision, so keep the scope narrow and start with a dry-run plan. The skill says scheduling, external posting and computer control need your explicit approval, and it covers credentials and private workspaces only briefly.
What are the alternatives to this skill?
If you want a simple one-off task, a normal Claude Code prompt without scheduling is enough. If you need a standalone agent framework instead of Claude Code, Hermes and AutoGPT are the frameworks this skill is designed to replace; this skill does not replace them outside Claude Code.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
Use this skill if you are a developer who wants Claude Code to run a recurring check, such as a daily pull request review, and you are comfortable reviewing scheduled prompts. Pick something else if you only need a one-off task, or if you handle sensitive accounts or private data and want stronger guarantees than the skill's brief notes provide.
Download the skill