Omnigent is an open-source orchestration layer that Databricks released in June 2026 under the Apache 2.0 licence. Harness Engineering: Anatomy, Architecture, and Evolution of Coding Agents — A Source-Code Study of Eleven Systems, which examines its source at version 0.4.0, describes it as the first Meta-Harness the authors are aware of: rather than implementing an agent loop of its own, it treats entire coding-agent harnesses — Claude Code, OpenAI Codex, Cursor, OpenCode, Hermes Agent, Pi and custom YAML-defined agents among them — as interchangeable components behind a common API.
The paper reads it as a bet that the Agent Harness has become a commodity component and that durable value sits one layer up, and uses it as a contrast point rather than as a corpus member: it implements no editing loop, no repository context and no edit-application strategy, so the authors decline to score it on the same subsystems as the harnesses they study.
Capabilities
On the paper’s account Omnigent runs as a four-tier process topology — server, host daemon, runner, and a per-conversation harness subprocess — whose adapter boundary is, by design, a recursive subset of its own public REST API. Its registry ships 23 canonical harness adapters plus aliases and a community entry-point group, formalized into five integration modes, and declared adapter capabilities are reconciled against a conformance bench that probes basic turns, tool calling, streaming, interrupts, model override and policy denial; the paper reports live verification for the four flagship SDK adapters and best-effort declarations for the rest.
Above that boundary the paper identifies four additions: composition, so that a Claude-brained orchestrator can dispatch work to Codex and have Cursor review it; a cross-harness policy plane with session, agent and admin levels, enforced on foreign harnesses through each vendor’s own extension mechanism such as Claude Code hooks, Cursor hooks, Hermes hooks and ACP permission requests, with fail-closed semantics; a uniform sandbox — bubblewrap plus seccomp on Linux, generated Seatbelt profiles on macOS and Job Objects on Windows — with a default-deny egress proxy hosting a credential proxy that keeps real tokens out of the sandbox; and server-durable sessions with multi-device sync, access grants, review comments, forking and mid-session harness switching.
The paper also notes what Omnigent does not do. Vendor-specific capability records and endpoints leak through its common API by design, and sandboxing is applied inconsistently — it wraps the Claude CLI in its own sandbox but delegates to Codex’s native sandbox modes — which the authors take as evidence that OS-level isolation resists being factored out as a shared service. Its two baseline dependencies are reported to be the claude-agent-sdk and openai-agents packages, which the paper uses to argue that harness SDKs have become the framework layer for a 2026 orchestration system.
Adoption & Ecosystem
The flagship example the paper describes, Polly, is a cross-vendor coordinator-worker setup: a Claude Code-brained orchestrator that writes no code itself, fans work out to six vendor harnesses in per-task git worktrees, and mandates in its prompt that the reviewer of a change come from a different vendor than its implementer. The paper uses Omnigent as evidence for its thesis that coding harnesses have turned into platforms, and lists the economics of such a layer — whether commoditization from above captures value durably — as an open question.