Deepagents-talon adds per-channel MCP authorization

LangChain's deepagents-talon scopes MCP server access by channel, per Version 0.0.7. That closes a permissions gap. Giving a Slack bot different tool access per channel used to mean running two separate agents, each with its own config.

Version 0.0.7 also hot-reloads MCP config without a restart. It adds OAuth plus device auth for the Slack and GitHub integrations.

Subagents gained new shapes too: background expendable instances, fork-mode for branching a task, fresh instances with no inherited state, and per-task tool selection.

If you're running a Slack agent against internal MCP servers, this addresses a risk you were likely already worried about. A finance channel talking to the same agent as engineering, with the same GitHub repo access, was a permissions problem waiting to surface. Per-channel authorization draws that boundary without forking the deployment.

The release notes don't say how the authorization gets enforced. A config flag the agent checks at call time is different from a boundary the MCP server itself refuses to cross. Only one of those survives a bug in the agent's own logic.