Unified Interface for Agent Harnesses

Everybody can connect an LLM to their backend in an afternoon but running a coding agent as a reliable product feature takes a lot longer.

This is because an agent needs a working directory, shell access, tools, permissions, session history, progress updates, artifact storage, cancellation, and a useful error when something breaks.

Now imagine your product supports both Codex and Claude Code. Each can complete sophisticated tasks, yet their CLI invocation, event formats and runtime assumptions differ.

Add Hermes or Pi and your integration layer starts to look like a pile of special cases.

In this case, we need a routing layer that exposes multiple agent harnesses through one task API, and ideally with an open specification such as “Unified Harness Protocol (UHP)”.

Let’s see how we can standardize the entire lifecycle of an agent task.

The Integration Tax Nobody Budgets For

Suppose you’re building an AI feature that investigates bugs, updates code and returns a patch. Your first prototype invokes a coding CLI and captures stdout.

Then users start asking reasonable questions:

• Can I see progress before the task finishes?
• Can I continue the same investigation tomorrow?
• Where did the generated files go?
• What happens if I click Cancel while a shell…