Claude Mods: Flipping the Agent Harness to Face the Human

Details

Claude Mods are TypeScript or JavaScript plugins that act as middleware for the Claude Code agent harness, allowing users to customize how the agent renders information and handles events like prompts or tool calls.
The system flips the traditional harness paradigm: instead of the agent interfacing with the machine, the harness becomes an 'Attention Interface' that the human and agent co-design to manage human attention and decision-making.
Anthropic ships six built-in mods, including 'You Should Know' (a side-agent that flags missed context) and 'diff', while users can create custom mods like 'Next Steps Supervisor' which suggests actions after turns with 8 or more tool calls.
Key limitations include noise (mods triggering on nearly every turn), cost (estimated 5–6k input-equivalent tokens per check on a 50k-token session), and security risks, as mods are not sandboxed and run with full user permissions.
The article contrasts Claude Mods with OpenAI's 'Intelligent UI' (model-designed interface) and Cursor's 'PR Routing & Approval' (team-policy driven), noting that Mods offer the most control but require users to manage the complexity themselves.

Key point