AI Agent Harnesses - spinningideas/resources GitHub Wiki

"Agentic Harness" Frameworks

Below is a collection of information about "AI Agent" orchestration systems that enable creating and running a "harness" that powers agentic development flows (or any effort that uses an AI Model that requires multi-step reasoning and tool use done in a structured way with some form of verification to get to "done")

TLDR

High Level TLDR:

Summary

A coding agent is two things: the model, and the harness wrapped around it. The model reasons. The harness gives it your context, your tools, and your process. One teardown of Claude Code found roughly 98% of it is the harness, not the model. The code analysis showed that the harness huge part of the "magic" of Claude Code and this is your true leverage - the model is becoming a commodity or can be substituted to some degree.

There are two levels to building a harness.

1) AI layer

The first level is a single coding agent session: the wrapper around one instance of Claude Code, Codex, or whatever you use.

The AI layer: your rules, skills, MCP servers, hooks, LSP, and subagents. Think BMAD or GitHub Spec Kit. Sounds a lot like context engineering, but what separates it is the mindset. When the agent does something dumb, you don't just blame the model - you improve the harness! It missed a convention, so that becomes a rule. It ran something destructive, so a hook now blocks it. Every mistake becomes a permanent upgrade to your system. See https://github.com/spinningideas/resources/wiki/AI-Assisted-Product-Development#general-ai-assisted-development-toolkits

2) Orchestration layer

The second level is orchestrating multiple coding agent sessions into one workflow (think Ralph loop - see https://github.com/spinningideas/resources/wiki/AI-Ralph-Technique). Given the context window restrictions of many models you should not hand a massive PRD to a single session and hope it can dissect it all. Instead, break it down into smaller, focused tasks and hand each task to a separate agent so that each agent has a fresh, clear, and focused context window to work with.

The orchestration flow breaks up the work in to multiple agents: One coordinates, one plans, one implements, one validates, and if everything passes the coordinator opens a pull request. Automate those handoffs and you get something like the Ralph loop, where the work is split into tasks and fresh sessions are run until the work is done and validated. That is how you take on larger tasks reliably without babysitting every step.

That second level is the real future of agentic engineering. The models and the tools keep getting better, and the way you scale alongside them is the harness you build around them.

Guides

High level guides and resources

Guides

Articles with general guides and high level overviews

Repositories

Various repositories with harness implementations and examples

Video Explainers

What is an Agent Harness?

An Agent Harness is the infrastructure that wraps around an AI model to manage long-running tasks. It is not the agent itself. It is the software system that governs how the agent operates, ensuring it remains reliable, efficient, and steerable.

It operates at a higher level than agent frameworks. While a framework provides the building blocks for tools or implements the agentic loop. The harness provides prompt presets, opinionated handling for tool calls, lifecycle hooks or ready-to-use capabilities like planning, filesystem access or sub-agent management. It is more than a framework, it comes with batteries included.

Agent Harness Architecture

We can visualize this by comparing it to a computer:

  • The Model is the CPU: It provides the raw processing power.
  • The Context Window is the RAM: It is the "limited", volatile working memory (eg the 256k token limit).
  • The Agent Harness is the Operating System: It curates the context, handles the "boot" sequence (prompts, hooks), and provides standard drivers (tool handling).
  • The Agent is the Application: It is the specific user logic running on top of the OS.

Agentic Harness Examples

The Agent harness implements "Context Engineering" strategies like reducing context via compaction, offloading state to storage, or isolating tasks into sub-agents.

Claude Code is a prime example of this emerging category, but there are a number of other frameworks and efforts emerging, and one could argue that all coding CLIs are, in a way, specialized agent harnesses given proper setup and a process calling them in a loop in a structured way.

Agent Harness Examples

Toolkits

Specific toolkits and libraries to help build harnesses

Additional Agentic "Frameworks"

There are a number of efforts that could also be considered some form of an "Agentic Harness" framework:

AI-Assisted-Product-Development

AI-Assisted-UI-Development

AI-Assisted-Development-Tools