The device agent can now report its own crashes back to the platform, with an admin switch to turn that off, and Claude sessions started from the unified Chat/Cowork mode are now captured the same way Claude Code sessions are. Kit assistants also got a round of reliability fixes to their built-in tools and underlying runtimes. More on this in Device Agent and Anthropic.

Features

• Device agents can report their own crashes #8171 - Device agents, including ones running on servers and CI that authenticate with an agent key, can now report fatal crashes, recovered panics, and update rollbacks to the control plane. Reports are scrubbed on arrival and forwarded to monitoring as logs and a crash count; they aren't stored. (Author: @bradcypert)
• Org admins can turn device agent crash reports off #8172 - Crash reporting is on by default. Organization admins can now turn it off for their fleet from the device agent configuration page. (Author: @bradcypert)
• The device agent can fetch your personal Claude observability plugin #8392 - Personal Claude accounts don't get the observability plugin installed automatically in Claude's cloud environment. The device agent now offers a download of your organization's plugin, keyed to your own account, for you to upload at claude.ai, so sessions in Claude's cloud environment are recorded under you instead of going unattributed. Each download replaces your previous one, and the shared org install key can't be used for it. (Author: @bradcypert)
• Claude Cowork sessions are captured the same way Claude Code sessions are #8445 - Claude sessions started from the unified Chat/Cowork mode are now captured: the observability plugin uploads new transcript lines whenever a turn finishes, and the platform stores them as the session's chat history. Policy enforcement stays specific to Claude Code; it doesn't apply inside the Cowork VM yet. (Author: @subomi)

Bug fixes

• Device agent crash reports stay apart by binary #8453 - Crash reports from the daemon and the menu bar app no longer get merged together when they share a function name, and reports can now carry the device's OS release. (Author: @bradcypert)
• Device coverage no longer borrows another machine's report #8557 - A device is no longer credited with a policy report from another machine unless that report names no other serial and comes from the same OS family, so, for example, a MacBook no longer shows "Enforcing policy" because the same person's Windows machine reported. (Author: @bradcypert)
• Kit assistants get built-in Speakeasy tools, more reliably #8459 #8475 - Kit assistants can now use built-in Speakeasy tools, including replying in Slack threads, alongside their attached MCP servers. A Kit runtime that can no longer serve turns is replaced automatically on the next message instead of refusing every turn, and slow sandbox startups no longer abort the runtime partway through. (Author: @danielkov)
• Kit picks up new conversations cleanly after a runtime swap #8430 - Kit assistants now reply to new Slack threads and to conversations that started before a runtime was replaced; a fresh session in a conversation with earlier turns is told that earlier messages are unavailable, instead of the turn failing outright. (Author: @danielkov)
• Retired Kit runtimes clean up promptly #8536 - Sandbox resources and workload credentials for a retired Kit runtime are now withdrawn within about an hour, even when the assistant stays busy. (Author: @danielkov)
• Kit assistants on GKE stop getting stuck mid-checkpoint #8440 - Idle Kit assistants on managed GKE now checkpoint and stop cleanly instead of staying stuck mid-checkpoint and refusing new turns. (Author: @danielkov)
• Kit can run without a separate authentication host #8447 - Kit assistants can now run on deployments without a separate authentication host; their workload issuer falls back to the server URL instead of refusing to start. (Author: @danielkov)
• The project home page loads faster #8684 - The project home page now makes fewer requests to load. (Author: @adaam2)