Claude Managed Agents get self-hosted sandbox execution

Claude Managed Agents can now run tool execution on self-hosted sandboxes instead of Anthropic's infrastructure, per the release notes. That's what matters for teams that need to audit what their agents actually did against production systems. Tool calls executed off your own servers are harder to lock down or review.

The same release wires Cowork into the Analytics API and OpenTelemetry, per the notes. That lets agent activity feed into whatever observability stack a team already runs.

Claude Code separately picked up sandbox credential blocking and org-level model restrictions, per the notes. Those are aimed at admins who need to control what agents can touch and which models they can call.

Claude Code also added a user-facing effort control, letting people choose how hard Claude thinks on a given response. That's a lever for cost and latency as much as quality, since not every reply needs full-depth reasoning.

Self-hosted sandboxes are the piece that matters most here. They let a team run agents against real infrastructure and still audit what those agents did. Cowork's OpenTelemetry tie-in only strengthens that case. All of it landed together in Anthropic's July update to Claude Code, alongside the effort control and admin restrictions above.