Introduction
Libra is a framework for building and operating long-horizon agent systems. It pairs a runtime Harness for composing agents with memory, tools, skills, sandboxes, and subagents, with an Application layer that turns those capabilities into a usable product.
If this is your first time with Libra, read the two pages in this section before going deeper into the Harness or Application docs. They establish the vocabulary the rest of the documentation assumes.
Why Libra exists
Most agent tooling optimizes for a single turn: take a prompt, call a model, return a response. Real research and engineering work is not single-turn. It spans many steps, accumulates context, runs code, reads results, and adjusts course.
Libra is built for that long-horizon shape. The architecture — a runtime harness underneath a product application — exists so that agents can carry state, act through tools and sandboxes, and stay coherent across an extended task instead of starting fresh on every message.
The two layers
- Harness — the Python SDK and runtime foundation. This is what you build on if you want to compose your own agent system: memory, tools, skills, sandboxed execution, subagents, and the middleware chain that orchestrates them.
- Application — a reference product built on top of the Harness. This is what you deploy and operate if you want Libra as a working application with a chat workspace, threads, artifacts, and configuration.
You can engage with Libra at either layer. Builders integrating Libra into their own systems live mostly in the Harness docs; teams deploying and running Libra live mostly in the Application docs.
What to read next
Once the concepts are clear, continue to the Harness docs to build on the runtime, or the Application docs to deploy and operate Libra.