September 30, 2026

Desktop 0.21.11

Big Salesforce work is rarely one job. It's a dozen jobs at once. With native subagents, one agent plans the work, spawns specialist agents to run each part in parallel, and brings their results back together. You can watch every one of them work in a new live agent tree.

What's New

Native subagents

Give one agent the goal, and it can now spawn specialist subagents to run parts of the work in parallel. For example, ask it to roll out what's new from Dreamforce 2026 in your org. One subagent migrates your agents to Agent Script, another turns on the new Data 360 features, and others roll out Lightning updates, update your Apex and set up permissions, all at the same time.

• Each subagent is a real agent with its own task, progress and results
• The lead agent coordinates the work and brings everything together when the subagents finish
• Subagents stay out of your sidebar, dashboard and search, so your agent list stays tidy

A live agent tree

See every subagent at a glance in a live tree above the chat. Each agent has its own animated glyph that shows what it's doing: starting, working, waiting for you, done. The tree updates as agents spawn and finish. Open any subagent to see exactly what it's working on.

Context model re-based to Claude Opus 5.5

The model underneath your agents is now Claude Opus 5.5, for sharper agents and smarter orchestration across long, multi-step work.

What's Improved

• Salesforce reconnects that pick up where you left off — reconnecting Salesforce updates your existing connection instead of creating a second one. If access has been revoked, MeshMesh now notices straight away and asks you to reconnect. Once you do, your agent resumes right where it left off, from whichever screen you reconnect on.
• Live progress checklist, back on every run — the step-by-step progress checklist shows up again as your agent works.
• Clearer notices when a connector needs you — if a connected tool (for example Linear or Notion) stops working, it's flagged for reconnect and you're told, instead of the agent quietly running without it.
• Quieter, safer agent output — context-management notices only appear when they actually affect your task, and sensitive values in tool output are masked more consistently.

What's Fixed

• Saving edits to a plan no longer fails intermittently
• Older desktop versions no longer show raw system messages in the chat

Security & reliability

• A high-severity dependency security patch
• Behind-the-scenes reliability improvements to support subagents at scale