LangChain’s Managed Deep Agents 0.8 introduces separate memory and credentials for each authenticated user, addressing a common problem when one production agent serves an entire organization. Personal preferences and recurring context can remain distinct from instructions shared by everyone using the deployment.
The runtime stores agent-wide and user-specific memory in separate mounted locations and does not copy content between them. By default, direct Slack messages can use both layers, while group conversations can use shared agent memory but not personal memory. Teams can change those policies for their own application.
Connections to services such as GitHub, Linear and Google Workspace can likewise use either a shared agent identity or the caller’s credentials. LangChain says 23 services receive built-in connection support. Version 0.8 also lets Slack users send files directly to an agent and adds HTTP channels for internal tools, customer portals and other systems that can deliver a JSON webhook.
A prebuilt web-search tool uses Parallel and records sources, latency and errors in LangSmith traces without requiring a separate vendor account. These managed features reduce infrastructure work, but developers still decide what an agent may remember, which credentials it can invoke and whether a group context should expose personal information. The safer defaults help only when those identity boundaries are preserved in custom tools.