The ways we contain Claude across products

Anthropic has developed containment strategies to safely deploy Claude across three agentic products—claude.ai, Claude Code, and Claude Cowork—by balancing the productivity gains of autonomous agents against their potential risks. The company has found that human-in-the-loop supervision is fallible (users approve ~93% of permission prompts, leading to approval fatigue), so it prioritizes technical containment measures like sandboxes and access controls to limit the "blast radius" of potential failures. As Claude's capabilities expand, the risk-reward calculation increasingly favors deployment when the agent's potential damage can be constrained through environmental boundaries, though some residual risk remains.