Changelog

What’s New

Approvals in the Dashboard Human-in-the-loop approvals now have a home in the dashboard. A new Approvals tab surfaces every pending request across your agents, so a reviewer can see what a tool is about to do and approve or reject it, and the run resumes from where it paused. The record captures who decided what, without ever storing the tool’s arguments or credentials. See Approvals.

Chaos Mode Test how your agents recover from failure before it happens in production. Chaos Mode injects bounded, deterministic tool and model faults into a single run, so you can confirm that fallbacks and retries behave as expected. Each run records exactly which faults were applied for later inspection. See Fallbacks and reliability.

Integrations on Tasks Integration tools can now attach directly to a Task, without wrapping it in an agent. A single task can call a connected service as part of its run, making lightweight automations simpler to build.

Tamper-Evident Audit Logs Audit logs now carry cryptographic integrity. Entries are chained so the log can be verified as unaltered, giving you a defensible, tamper-evident record for compliance. See Observability and audit.

Improvements

• Task tools attached to an agent can follow the task’s active revision, so improving a task rolls out to every agent that uses it. See Versioning and revisions
• Richer Slack responses with formatted text, tables, facts, and sources, plus live progress while the agent works
• Gemini context caching lowers the cost of runs that reuse the same context
• Agent Observe adds server-side run-history filters and surfaces primary-model and integration failures with specific reasons
• Tags can now carry a description and be updated in place