Already 3 months old when the scout caught it.
Meta-harness
What it is
A meta-harness is a control layer that sits above multiple individual agent harnesses — Claude Code, Codex, Cursor, custom agents — rather than living inside or replacing any single one of them. It lets you compose several coding agents into a team, apply shared policy (cost budgets, sandboxing, approval gates) across all of them uniformly, and keep session memory, tool configuration, and skills portable even as you swap which underlying harness or model actually does the work.
Why it's emerging now
As teams routinely run several different agent harnesses side by side — Claude Code, Codex, Cursor, Pi, homegrown agents — the friction has shifted from "does any one agent work well" to "how do I coordinate, govern, and share work across all of them at once." Databricks named this gap explicitly, calling Omnigent a "harness of harnesses" when it open-sourced the project, and the framing was picked up almost immediately by other teams building the same kind of layer for their own tools (notably Nx's Polygraph), which suggests it names a real architectural pattern rather than one vendor's marketing term.
How it relates to what you know
It builds on harness engineering (the idea that "Agent = Model + Harness") by adding one more level of indirection: harness engineering is about designing the scaffolding around a single model, while a meta-harness designs the scaffolding around several harnesses at once. It's a different axis from managed agents, which is about hosting one agent's runtime in the cloud — a meta-harness can sit above managed and local agents alike, and its job is cross-harness coordination and policy, not hosting.