When we talk about AI agents, we usually focus on the model, prompts, and tool calls. But there's a lot more happening behind the scenes.
Context engineering, memory, guardrails, policies, approvals, execution, tracing, evaluations, and releases are all part of building an agent system. Yet these pieces are rarely explored together in a way that's easy to understand.
I wanted to see these pieces work, not just read about them.
So I built Harnesswalk, an open-source, interactive environment for exploring how an AI agent harness works, one node at a time.
What is Harnesswalk?
Think of it as a visual playground for understanding AI agent engineering.
Instead of looking at a static architecture diagram, you can interact with an agent's execution flow and see what happens at each stage.
Some things you can explore:
- Step-by-step execution: Follow a request as it travels through the harness, one node at a time.
- Prompt inspection: See the exact prompt being prepared before a model call.
- Component inspection: Explore the code and configuration behind individual nodes.
- Human-in-the-loop: Approve or reject tool actions before execution.
- Event replay: Go back and inspect what happened during an execution.
- Experimentation: Disable individual nodes and observe how the system behaves without them.
The idea is to make the architecture tangible rather than treating the agent as a black box.
A note on scope
Harnesswalk currently models a governed assistant with policies, approvals, audit, memory, and a release gate.
It's a teaching model, not a debugger for existing agents or a coding-agent harness with shell access and subagents.
It's still evolving, and there are plenty of things I'd like to improve. But the webapp is already available, runs in the browser without installation, and supports an offline mode using a scripted model.
What's next?
I'm interested in exploring how this could evolve beyond an interactive learning environment.
Some directions I'm considering include multimodal agent workflows, more extensible component contracts, and eventually a registry where developers can add or replace capabilities such as RAG, memory implementations, model providers, and evaluation workflows.
For now, I'm focusing on making the existing experience useful and learning from people building agentic systems.
Give it a try
🌐 Explore the Harnesswalk webapp
📝 A deeper look at the architecture
It's open source under the Apache-2.0 license.
I'd love to hear from other developers:
What's missing from the architecture? Which components would you add, replace, or remove? And how would you approach making an agent harness more observable and easier to understand?
Feedback, ideas, and contributions are welcome!
Top comments (0)