HKUDS/OpenHarness
OpenHarness
OpenHarness is a framework for building and orchestrating AI agents that utilize tools and plugins to execute complex tasks. It provides an orchestration system for managing language model lifecycles and a multi-agent coordination system for delegating workloads across teams of specialized subagents.
The project features an agent gateway that bridges language model agents to external chat platforms and communication channels. It includes a tool integration engine for executing shell, file, and web operations, supported by a memory and skill manager that handles persistent user preferences and the on-demand injection of domain-specific knowledge.
The framework incorporates security via path-based access controls and interactive approval workflows for sensitive operations. It also provides a dry-run mode for execution plan inspection, a plugin architecture for custom lifecycle hooks, and a terminal user interface for session management and real-time feedback.
Features
- Agentic LLM Frameworks - Provides a platform for building and managing autonomous agents powered by LLMs with tool use and memory support.
- AI Agent Orchestration Frameworks - Provides a framework for managing autonomous agent loops and tool discovery to execute complex tasks.
- Multi-Agent Coordination Systems - Implements a delegation system allowing specialized agents to collaborate on complex tasks by sharing state.
- Recursive Subagent Nesting - Manages complex workflows via recursive delegation to specialized subagents in independent lifecycles.
- Communication Gateways - Ships a gateway that routes messages between external communication platforms and AI agents.
- Agent Orchestrators - Provides a system to manage the reasoning loops and execution lifecycles of language model-based agents.
- Agent Memory Managers - Provides a system for managing persistent user preferences and the storage and retrieval of agent memory.
- LLM Tooling Integrations - Connects language models to shell, web, and file system tools with type-safe validation.
- Custom AI Assistant Development - Enables building personalized AI assistants with custom plugins, on-demand skills, and persistent memory.
- Multi-Agent Orchestration Systems - Orchestrates teams of specialized autonomous agents to execute complex, collaborative workflows.
- Tool Execution Permissions - Enforces security through path-level access rules and interactive approvals for sensitive tool executions.
- Agent Gateways - Provides a gateway that bridges AI agents to external chat platforms and communication channels.
- Dynamic Skill Injection - Implements on-demand injection of domain-specific markdown knowledge into the agent's context.
- LLM Provider Integrations - Offers a configurable profile system for managing connectivity and authentication to various AI provider backends.
- Type-Safe Execution Frameworks - Ships a type-safe validation system for executing shell, file, and web tools.
- Agentic Memory Systems - Ships a storage system for maintaining long-term user preferences and interaction continuity.
- Tool Execution Inspection - Includes a dry-run mode for inspecting the intended tool calls and actions before they are executed.
- Agent Session Persistence - Maintains continuity across interactions by storing user preferences and knowledge in dedicated local files.
- Path Access Restrictions - Restricts agent file system operations using path-based security rules and interactive user approvals.
- Data Type Validation - Ensures language model outputs match expected schemas through strict type validation before tool execution.
- AI Backend Abstractions - Decouples core logic from specific models through a provider-based abstraction for remote and local backends.
- Extensible Plugin Architectures - Provides a modular extension system allowing custom commands and lifecycle hooks to modify agent behavior.
- Skill Files - Implements modular, domain-specific instruction files loaded into the agent context to guide task execution.
