Claude Mods are plugins for Claude Code written as TypeScript or JavaScript functions that run inside the CLI and desktop app. They respond to events such as prompts, tool calls, and interface renders, and can observe, modify, or override default behavior. The article describes them as middleware for the agent harness, injecting code that changes how Claude Code executes and renders at runtime.
Dan McAteer argues this marks a reversal in the harness's direction. Originally a layer between the agent and the machine, the harness is becoming a layer between the agent and the human: an "Attention Interface" that decides when to interrupt, what to show, and which decisions require the user. Because users can edit or create mods, and can ask Claude to build mods in plain language, the human-facing part of the harness becomes programmable by both sides.
Anthropic already ships six built-in mods, including AGENTS.md loading, diff rendering, plugin authoring, security defaults, telemetry, and "You Should Know." The author highlights You Should Know because it runs a side agent that flags material details hidden in the agent trace. In his example, it caught uncommitted work in roughly 45 files before a discard would have destroyed it. He also built "Next Steps Supervisor" as part of the open-source ModSmith plugin, which suggests a clickable next action after substantial turns.
With only one source, there are no differing accounts to compare. The article points to OpenAI's Intelligent UI and Cursor's PR Routing & Approval as similar efforts elsewhere, and makes the broader bet that the harness becomes a platform for human-agent interfaces, with models as the engine.