Agent Meta-Harness

The Agent Meta-Harness is the orchestration engine at the heart of Voir. Where an Agent is a declaration of intent and scope, the Meta-Harness is the machinery that, on every single chat turn, re-assembles that declaration into a running agent — fetching tools from live connectors, wrapping them in governance guards, composing the system prompt, retrieving relevant knowledge, and streaming a token-and-tool event loop back to the caller.

It is called a meta-harness because it does not itself answer questions. It builds the thing that answers questions, and it does so with three non-negotiable properties:

• Fault isolation — one dead or slow connector must never take down the whole agent. Tools are fetched per-connector with independent timeouts; a failure degrades gracefully.
• Enforce-in-code governance — the harness wraps every scope-sensitive tool so that document-partition and SQL-allowlist boundaries are enforced inside the tool call, before it reaches any downstream system. The model's cooperation is never assumed.
• Progressive disclosure — rather than pay for a giant prompt every turn, the harness gives the agent compact context plus meta-tools it can call to pull in detail (skills, full scope enumerations) only when needed.