Unlock Production‑Grade AI Agents with the OpenHarness Python Framework

OpenHarness Overview

OpenHarness is an open‑source Python implementation of the Harness agent framework. It provides a lightweight, production‑grade infrastructure for building AI agents that can use tools, maintain contextual memory, enforce permission policies, and coordinate multiple agents.

Key Capabilities

Agent Loop – Streams model responses, detects tool calls, retries with exponential back‑off, executes tools in parallel, and tracks token usage.

Toolbox – Ships with more than 40 built‑in tools (file I/O, shell commands, web search, MCP protocol, etc.) and supports on‑demand loading of skill definitions from Markdown files.

Context Memory – Automatically injects a CLAUDE.md system prompt, compresses long contexts, persists knowledge in MEMORY.md, and restores sessions across runs.

Permission Governance – Multi‑level permission modes (default, automatic, scheduled), path‑based allow/deny rules, and interactive approval hooks for safe tool execution.

Group Coordination – Generates sub‑agents, maintains a team registry, and manages the lifecycle of background tasks for collaborative workflows.

Repository

https://github.com/HKUDS/OpenHarness

Core Architecture

The repository is organized into ten subsystems that together implement the agent framework:

openharness/
  engine/          # Agent loop – query → stream → tool call → loop
  tools/           # 40+ tools – file I/O, shell, search, web, MCP
  skills/          # Knowledge base – load skill Markdown files on demand
  plugins/         # Extension points – commands, hooks, custom agents, MCP server
  permissions/     # Security – multi‑level modes, path rules, command denial
  hooks/           # Lifecycle – pre/post tool‑use event hooks
  commands/        # 54 built‑in commands (e.g., /help, /commit, /plan, /resume)
  mcp/             # Model Context Protocol client implementation
  memory/          # Persistent cross‑session memory store
  tasks/           # Background task scheduler and manager
  coordinator/     # Multi‑agent coordination – sub‑agent creation, team registry
  prompts/         # System prompt assembly, CLAUDE.md injection, skill integration
  config/          # Hierarchical configuration and migration utilities
  ui/              # React‑based terminal UI (backend protocol + frontend)

Core Loop Logic

The main execution loop continuously streams model output, processes tool calls, and feeds results back to the model:

while True:
    response = await api.stream(messages, tools)

    if response.stop_reason != "tool_use":
        break  # Model finished its reasoning

    for tool_call in response.tool_uses:
        # Permission check → pre‑hook → tool execution → post‑hook → result
        result = await harness.execute_tool(tool_call)
        messages.append(result)

    # Continue the loop so the model can act on the new messages

Getting Started

After cloning the repository, the entire framework can be launched with a single command: oh This command starts the OpenHarness server and loads all default components, allowing immediate interaction with the agent via the built‑in CLI or programmatic API.