A meta-harness, in the usage of Harness Engineering: Anatomy, Architecture, and Evolution of Coding Agents — A Source-Code Study of Eleven Systems, is an orchestration layer that coordinates one or more coding-agent harnesses from above and implements no editing loop of its own. Where an Agent Harness wraps a model to make it act on a repository, a meta-harness wraps entire harnesses — Claude Code, Codex, Cursor and others — as interchangeable components behind a common interface. The paper treats it as one of four things a harness is not, and scoring it on the same subsystems as a harness would, in its words, be a category error; it analyses the first such system it is aware of, Omnigent, as evidence about which harness capabilities are migrating up into a separate layer.
Usage
The paper describes what the meta-layer adds above the harness boundary in Omnigent’s case: composition, so that any registered harness can be addressed as a sub-agent session of any other; a cross-harness policy plane enforced on foreign harnesses through each vendor’s own extension mechanism, such as Claude Code hooks or ACP permission requests; a uniform sandbox and egress stack; and shareable, server-durable sessions that can switch harness mid-session. What stays below the line is the harness core — the editing loop, repository context and edit-application strategy.
The authors read the meta-harness as a bet that the harness is becoming a commodity component and that durable value sits one layer up, and they place it within their broader thesis that coding harnesses have turned from tools into platforms. They note two tensions in that reading from the one system they examine: the meta-layer does not pretend the harnesses are equivalent, with vendor-specific capability records leaking through its common API by design, and it applies sandboxing inconsistently — wrapping one vendor’s CLI in its own sandbox while delegating to another’s native sandbox modes — which the paper takes as evidence that OS-level isolation resists being factored out as a shared service. It lists the economics of commoditization from above as an open question for future work.
The paper’s analysis rests on a single system and a single source-code study; it calls Omnigent the first meta-harness it is aware of rather than a representative of an established category.
Anthropic uses the same word for a different design. In Scaling Managed Agents: Decoupling the brain from the hands it calls Claude Managed Agents “a meta-harness”, meaning a system that is unopinionated about the specific harness Claude will need in future and instead offers general interfaces — a durable session and sandboxes reached through a tool-call interface — that many different harnesses can run on, from a general-purpose one such as Claude Code to task-specific ones. Where the paper’s meta-harness coordinates existing vendors’ harnesses from above, Anthropic’s hosts whatever harness sits between Claude and those interfaces, and it describes the design as being opinionated about the interfaces around Claude rather than about the harness itself (see Brain/Hands/Session Split). Neither source refers to the other’s usage.