Multi-harness
Also known as: harness-agnostic, cross-harness, multi-harness orchestration
An agent harness is the infrastructure layer that wraps a coding model: it handles tool calls, sandboxing (isolated execution environments), permissions, context management, and recovery logic. For most of 2025 and early 2026, teams picked one harness and committed to it. Multi-harness emerged as a pattern when enterprises realized different harnesses have different strengths and they didn't want to be stuck with one.
In a multi-harness setup, a control plane sits above the individual harnesses and decides which one to use for a given task, sometimes running several in parallel. Claude Code might handle complex reasoning tasks, Codex handles certain language-specific jobs, and Warp Agent handles terminal operations, all coordinated through one interface. Persistent memory and skills (reusable agent behaviors) can be shared across harnesses so knowledge doesn't silo.
The wshobson/agents repo and Warp Oz are examples of multi-harness approaches gaining traction in October 2026. The architectural question it answers: how do you write your agent skills, prompts, and memory once and deploy them across whichever harness the team is running this week? It is a specialist concern for now, but the underlying idea of harness portability is likely to become a standard enterprise requirement as agent infrastructure matures.