Harness Studio — See Why Your AI Agent Failed

When an AI agent fails, you usually only see the wrong answer — not that it looped, lost the context it needed, or called a tool it shouldn't have. Without visibility into each step, fixing an agent is guesswork.

A minimal agent harness in Python, written from scratch — the layer that sits between the model and the outside world:

Agent loop that plans, calls tools and decides when it is done

Context compaction so long runs stay within the model's context window

Permissioned tools — an agent can only call the tools it has been granted

Sub-agents that take on delegated tasks and report back

Harness Studio, a web app that replays every run:

One lane per agent, so you can follow the main agent and its sub-agents side by side

Each step explained — what the agent decided, which tool it called and what came back

Exactly what each agent still had in context when it stopped

This is the same thinking I bring to client agent work: build the agent so its behaviour can be inspected, limit what it is allowed to do, and test it — so problems are found before your users find them.

Stack: Python, JavaScript, GitHub Actions (tests and Pages deployment)

Live demo: https://rolinmuuu.github.io/Mini-Agent-Harness/ Code: https://github.com/Rolinmuuu/Mini-Agent-Harness