Claude Code is becoming more customizable through Mods, a system that lets users alter the agent harness around the model. In a Latent Space interview, Anthropic’s Thariq Shihipar described the feature alongside Projects, cloud sessions, plugins and collaborative Claude Tag workflows.
The harness controls how a coding agent gathers context, uses tools and structures work. Opening it to Mods gives advanced users more influence over that process without changing the underlying model. Anthropic is also developing a split between a cloud-based “brain,” local or remote “hands,” and interfaces generated for the current task. Persistent artifacts can become shared work surfaces rather than one-off outputs.
Customization increases flexibility, but it can also weaken assumptions made by the default safety design. A modified harness may introduce new tools, permissions or communication paths that the model can misuse. Shihipar’s discussion frames sandboxing, prompt injection, probes and fallback systems as core engineering concerns. Teams adopting Mods should treat community extensions like executable infrastructure: inspect them, restrict credentials and test behavior before use on real repositories.