Agents that improve themselves

Automatic improvement, on your terms

Agents can now propose improved versions of their own instructions and working steps. Preview mode lets you review every suggestion before it goes live; On mode applies reviewed improvements automatically; Off keeps things exactly as you set them.

Clearer approvals

Approval requests are now separated from run activity and spelled out in plain language, so it's obvious what you're agreeing to before an agent acts.

Improvements

• Runs blocked by an expired connection say so up front and link straight to reconnecting.
• Set a default thinking level per model, so each agent reasons as deeply as the job needs.
• You can point runs at an optional help destination and resume interrupted requests.
• New connector: Tella.
• The app does noticeably less work behind the scenes — pages load faster and live updates arrive sooner.

Fixes

• Disconnecting an app from the connector menu works reliably again.
• The agent detail header no longer crowds on smaller screens.