ToolExperimentalA2AMCP

harness

Name: harness
Rating: 3.0 (40 reviews)
Author: mastersof-ai

by mastersof-ai

TypeScript agent runtime with Markdown-defined agents and TUI/web UIs

TypeScript

Updated Mar 19, 2026

Stars

Forks

View on GitHub

What It Does

Provides a TypeScript agent runtime with full system-prompt control and programmable agents defined in Markdown. Runs agents locally in a terminal TUI or a multi-user web UI and is powered by the Claude Agent SDK for agent execution. Exposes agent definitions, interaction flows, and runtime controls so teams can iterate on prompts, roles, and handoffs quickly, leveraging A2A Protocol Pattern and Tool Use Pattern.

Key Benefits

As agents are asked to delegate and interact, reproducible system-prompt control and readable agent definitions become essential for auditing and trust. Harness makes it easy to inspect how agents are wired, replay conversations, and standardize prompt/role artifacts—helping teams surface failure modes and build agent track records. For evaluation workflows, having a runtime that preserves prompt state and multi-user views lowers the friction to run A2A evaluation and collect behavioral signals. This supports stronger governance with Human-in-the-Loop.

When to Use

Developers prototyping multi-agent flows who want fine-grained system prompt control and an inspectable runtime. This toolkit fits teams building and testing orchestrations, where you might rely on a registry-like pattern such as the Agent Registry Pattern to manage agent definitions and capabilities.

Real-World Examples

Iterate on and version system prompts and agent roles using Markdown artifacts
Replay multi-agent conversations in a TUI or web UI to diagnose failure modes
Prototype agent delegation patterns and collect interaction logs for A2A evaluation

Need implementation help?

Expert guidance available