OpenRig Multi-Agent Harness Runs Claude Code and Codex Together

What Happened

OpenRig, a multi-agent harness published on GitHub under the repository mvschwarz/openrig, orchestrates Anthropic's Claude Code and OpenAI's Codex as a single coordinated system rather than as isolated tools invoked in sequence. The repository gained 114 stars in a single day. The project positions itself as a cross-vendor runtime layer, meaning agent dispatch, context handoff, and task decomposition are managed by the harness rather than by either vendor's CLI independently.

Why It Matters

Coding-agent workflows have largely been single-vendor: a team standardizes on Claude Code or Codex and builds prompt scaffolding, context handling, and review loops around that one tool. OpenRig treats the model and CLI as interchangeable workers behind a common orchestration surface, which lets teams route subtasks by strength — for example, one agent for planning and another for patch generation or verification. The immediate beneficiary is the platform or tooling team that wants to A/B different coding agents inside the same pipeline without rewriting glue code. It also creates an early reference point for how cross-vendor agent interop might be standardized at the harness layer rather than at the model provider layer.

Technical Details

OpenRig wraps both Claude Code and Codex as interchangeable agents inside a single harness, exposing a shared task interface rather than vendor-specific invocation. The repository is hosted on GitHub at mvschwarz/openrig and is early-stage; the 114-star single-day velocity is the primary signal of attention rather than any published benchmark. The project is a harness — it manages agent lifecycle, task routing, and context passing — not a model or a training contribution. Integration depends on the user having credentials and CLI access for both vendor tools, which means the harness inherits each vendor's rate limits, pricing, and authentication constraints. No public performance numbers, benchmark comparisons, or formal capability claims accompany the release.

Operational Impact

For teams already running multi-step coding pipelines, the practical change is that agent selection becomes a routing decision rather than an architectural commitment. Context handoff between agents — a task that today is usually a brittle combination of file writes, prompt re-injection, and manual copying — becomes a harness responsibility, which reduces per-workflow glue code. Cost accounting shifts: operators can direct cheap tasks to one vendor and expensive reasoning to another, then measure per-agent spend in one place. The friction this removes is small per-developer but compounds across teams that previously maintained parallel Claude Code and Codex wrappers. It also creates a new dependency surface — the harness must track both vendors' breaking changes to CLI flags, auth, and output formats.

What To Watch

The signal to track is whether cross-vendor harnesses stabilize into a shared interface convention — comparable to how LangChain or OpenAI-compatible APIs converged — or fragment into per-harness dialects. If OpenRig's star velocity translates into maintained releases, expect adjacent projects to standardize agent handoff protocols and to add support for other vendors (Gemini, local models) as dropout replacements. The second-order effect is on vendor lock-in: if orchestration lives above the model, the competitive moat shifts from the CLI to the routing and evaluation layer.