Cursor 3.11: Side Chats, Team MCP, and Cloud Agent Hooks

Cursor 3.0 handed you a fleet of parallel AI agents in April. Cursor 3.11, released July 10, is the update that makes running that fleet something other than chaos. It ships four improvements your agent workflow has needed: side chats so you can ask questions without interrupting a running agent, searchable conversation history so you can find what any agent did last Tuesday, centralized MCP server distribution so your whole team and your cloud agents share the same configuration, and five new lifecycle hooks for cloud agents — though the last one comes with a production caveat you should hear before you build on it.

Side Chats: Ask Questions Without Stopping Your Agent

Here is the problem side chats solve. Your main agent is mid-task implementing auth middleware. You need to know whether to use jose or jsonwebtoken. Your options before Cursor 3.11: interrupt the agent with your question and reset its context, or open a second full chat window and lose the connection to what the main agent is doing. Most developers learned to just hold the question and wait.

Side chats are a third option. Type /side or /btw in the chat input — or click the plus button at the top of the panel — and you get a secondary conversation running alongside the main one. By default, side chats are scoped to reading and analysis only. They will not start making changes while your main agent is already touching files. Ask your question, get your answer, then @-mention the side chat from the main thread to pull that conclusion back in. The main agent never stopped.

Each side chat is a full durable conversation — you can follow up, revisit it from the Agents Window later, and @-mention it in any future main thread. This is not a throwaway dialog. It is a permanent secondary agent session with its own history, which matters once you start using side chats regularly and want to trace a decision back to where it came from.

Agent Transcript Search: Find What Your Fleet Did

If your team has been running parallel agent sessions since Cursor 3.0, you have already hit this: something broke, and you need to find which agent touched that file three sessions ago. Before 3.11, finding it meant scrolling. Some developers built workarounds using the open-source ctxrs/ctx project just to get basic search over agent history. That workaround is now unnecessary.

Cursor 3.11 ships a local search index over all agent conversations. In the Agents Window, Cmd+K opens cross-session search against actual message content — not just conversation names or PR numbers. Within a single conversation, Cmd+F gives you match-jumping and a match counter. The index builds locally, which means your agent transcript content does not route through Cursor’s servers to be indexed. For teams with sensitivity about what their agents produce, that distinction matters.

Team MCP Distribution: End Configuration Drift

MCP server configuration was per-developer before 3.11. Every engineer on your team configured their own connections to internal APIs, Figma, Linear, or whatever authenticated tools your agents needed. Cloud agents had no credentials at all. The result: configuration drift, inconsistent tool access between team members, and cloud agent runs that failed because nobody had provisioned credentials for the VM.

In 3.11, admins configure Team MCP servers once under Dashboard > Integrations & MCP. Those servers propagate automatically to cloud agents, the Agents Window, the local IDE, and the CLI. A Team Marketplace lets admins push approved integrations for one-click install — team members get consistent access without manual configuration. Enterprise plans can scope marketplace access to specific SCIM directory groups, which is useful for separating what contractors can access from what full-time engineers can. If your team runs cloud agents against any authenticated service, this setup is worth doing today. The reduction in “it works on my machine” agent failures alone justifies it.

Cloud Agent Hooks: Promising, Not Production-Ready

Cursor 3.11 adds five new lifecycle events to the hook system in .cursor/hooks.json: beforeSubmitPrompt, afterAgentResponse, afterAgentThought, subagentStart, and stop. The vision is self-correcting agent loops — your hook checks the agent’s output quality after each response and queues a follow-up prompt if it fails. Compliance audit trails, policy enforcement, PII stripping before prompts reach the model: hooks are what connect cloud agents to production-grade operations.

Do not build production workflows on afterAgentResponse or stop yet. Community bug reports on Cursor’s forum confirm these hooks do not reliably fire in cloud agent VMs as of 3.11. There is a compounding problem: the beforeSubmitPrompt hook fires correctly and returns a modified prompt, but the modification is silently stripped — the model never receives it. Cursor has acknowledged both issues; a fix is expected in a forthcoming patch. What works reliably today: file and tool hooks (preToolUse, beforeShellExecution, afterFileEdit) and afterAgentThought. Stay within those if you need hook-based automation in production now.

What to Do Now

• Upgrade to Cursor 3.11 — side chats and transcript search are available immediately after update.
• Configure Team MCP servers at Dashboard > Integrations & MCP if your team runs cloud agents against authenticated services.
• Start using /side or /btw in your next agent session. It changes how you think about mid-task questions.
• Hold off on building production workflows around afterAgentResponse or stop hooks until Cursor patches cloud VM hook execution.
• Watch the Cursor changelog for the 3.12 patch. Hooks are the foundation of agent self-correction — once they work reliably, the automation possibilities open up considerably.

Cursor 3.0 was the bet that developers want agent fleets, not just a smarter autocomplete. Cursor 3.11 is the acknowledgment that running a fleet requires a management layer. Side chats and transcript search are the pieces that should have shipped in April. Better late than broken.