AgnesAI-Labs/agnes-harness

What it solves

Agnes Harness (AGH) addresses the challenge of deploying AI agents into complex, real-world enterprise environments. It provides a framework to integrate custom business systems, internal APIs, and physical devices into an agent's workflow, ensuring that every action is recorded, auditable, and subject to human approval.

How it works

The project acts as a unified runtime foundation with four architectural pillars: the Brain (LLM for reasoning), the Cerebellum (Jev for structured decision-making), the Memory (Harness for session history and task state), and the Body (MHS for device interaction). Developers build plugins to connect business-specific tools, which the agent can then execute within a sandboxed environment. The system maintains a shared state across CLI, Web, and SDK interfaces, allowing users to start tasks in a browser and continue them in a terminal without losing context.

Who it’s for

It is designed for Forward Deployed Engineers (FDE) and solution developers who need to deliver AI agents that interact with proprietary customer systems, require strict oversight, or need to coordinate physical hardware through standardized protocols.

Highlights

• Unified Session Management: Seamlessly switch between Web, CLI, and SDK interfaces while maintaining full task history and state.
• Auditability and Control: Features a Trace view that records every model call, tool execution, and human approval, ensuring transparency in agent behavior.
• Pluggable Architecture: Allows for the creation of backend tools, frontend panels, and full-stack plugins to extend agent capabilities for specific business needs.
• MHS Protocol: Includes a dedicated Model Hardware Standard (MHS) to integrate physical devices and sensors into agent workflows, independent of vendor-specific SDKs.