A structured workflow framework based on Anthropic engineering principles to manage complex AI agent projects across multiple context windows and sessions.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install long-running-harness
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install long-running-harness using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is Long-Running Agent Harness?
The Long-Running Agent Harness is a technical framework designed for Openclaw Skills to handle tasks that exceed the limitations of a single context window. Based on the Effective Harnesses for Long-Running Agents methodology by the Anthropic engineering team, it prioritizes persistent file-system state over volatile agent memory.
By utilizing structured JSON for state tracking and Markdown for logging, this skill ensures that AI agents can resume work reliably, maintain focus on incremental goals, and verify progress through rigorous testing. It transforms Openclaw Skills into robust project managers capable of handling software development, research, and complex data analysis without losing context or quality.
Long-Running Agent Harness Use Cases
- Initializing new long-term projects with standardized directory structures.
- Managing complex software development cycles involving dozens of incremental features.
- Resuming development or research work after a context window reset or session break.
- Orchestrating sub-agents for specific tasks while maintaining global project context.
- Generating automated progress reports and project status summaries for stakeholders.
How Long-Running Agent Harness Works
- Project Initialization: The agent creates a standardized project structure including metadata files like PROJECT.md and features.json.
- Task Mapping: High-level goals are broken down into granular, testable features within a structured JSON schema.
- Session Startup: At the beginning of each session, the agent reads the progress log, git history, and feature list to synchronize its state.
- Incremental Execution: The agent focuses on a single feature per session to ensure focus and prevent scope creep.
- Verification: Every change is validated against predefined tests or manual checks before being marked as complete.
- State Persistence: The agent updates the feature tracker and progress log, then commits all changes to Git before ending the session.
Long-Running Agent Harness Setup
To start a project using this framework within Openclaw Skills, initialize your project directory and create the required metadata files:
# Create project directory
mkdir -p projects/<project-name>
# Initialize git repository
git init
# Create core tracking files
touch projects/<project-name>/PROJECT.md
touch projects/<project-name>/features.json
touch projects/<project-name>/progress.md
Trigger the skill by saying "Start project" or "Initialize workflow" followed by your project details.
Long-Running Agent Harness Data Schema & Taxonomy
The harness organizes data using a strict file-based taxonomy to ensure persistence:
| File | Purpose | Format |
|---|---|---|
PROJECT.md |
Defines project goals, tech stack, and acceptance criteria. | Markdown |
features.json |
A structured list of tasks with priority, status, and test steps. | JSON |
progress.md |
An append-only log recording every session's activity and hurdles. | Markdown |
init.sh |
Optional script for environment and dependency setup. | Bash |
src/ & tests/ |
Source code and corresponding validation suites. | Directory |
Long-Running Agent Harness Advanced Features
- Sub-Agent Dispatch: Uses the sessions_spawn protocol to delegate specific features to child agents with self-contained context packages.
- Cron Job Auditing: Supports scheduled turns to inspect progress, detect bottlenecks, and generate status alerts.
- Automated Git Integration: Every successful feature implementation is coupled with a structured git commit for easy rollbacks.
- Failure Mode Prevention: Hard constraints prevent the agent from marking tasks as passed without explicit verification steps.
- Domain Agnostic: Can be extended from coding to research, technical writing, or data analysis by modifying the feature category taxonomy.
SKILL.md
Loading
Related Openclaw Skills
CrewAI Team Collaboration
A sophisticated multi-agent orchestration skill that leverages CrewAI to transform raw product ideas into professional Product Requirement Documents.
Interview Question Generator & Evaluator
A specialized tool for generating structured Feishu interview documents from resumes and providing comprehensive assessments from transcripts.
QMAP Client
A powerful CLI tool designed for interacting with the QuantMap distributed computing protocol to manage node operations and task execution.
Shuidichou Assistant
A specialized assistant for China's largest medical crowdfunding platform, enabling real-time donations and streamlined fundraising support.
mcp-to-skill
An automation tool that transforms Model Context Protocol (MCP) servers into standalone, zero-dependency skill packages for AI agents.
ITSM Ticket Automation Skill
A lightweight automation tool for submitting enterprise ITSM tickets using Python and system-level Chromium.
Featured*
