LangChain released version 0.8 of Managed Deep Agents on September 24, adding a second layer of memory that lets a deployed AI agent keep private notes on each person who talks to it, separate from what it remembers about the team as a whole. The release also adds new ways to reach an agent from outside Slack and folds in a web search tool at no extra cost while the service stays in beta.
Managed Deep Agents is LangChain’s hosted wrapper around Deep Agents, its open source toolkit for building AI systems that plan and act over long stretches rather than answering a single prompt. The wrapper handles the plumbing that turns a working prototype into something a company can run in production: logins, permissions, memory storage, and the channels people use to talk to the agent. LangChain says building that plumbing from scratch typically consumes a company’s engineering roadmap for a quarter or more.
The new per user memory solves a specific trust problem. When one agent serves an entire team, mixing up whose preferences belong to whom is embarrassing at best and a data leak at worst. LangChain’s fix stores team wide instructions in one folder and each person’s private context in another, and the two never copy into each other unless a developer wires them together on purpose. A support agent can hold shared escalation rules in the team folder while remembering that one teammate prefers short Slack updates in their own.
“Our production chat app runs as a Managed Deep Agent. With Context Hub and user level memory, we are able to maintain client confidentiality while delivering a superior experience,” said Zahid, chief technology officer at Kyth.ai, in a testimonial LangChain published alongside the release.
Credentials get the same split. An agent can hold its own shared login for a tool everyone uses the same way, such as web search, while pulling a specific person’s own credentials for tools like GitHub or Salesforce where permissions differ user to user. LangChain says the platform now ships with ready made connections for 23 services, including Linear, GitHub and Google Workspace tools, so a developer only has to supply a user ID and a secret rather than build the authorization flow.
On the access side, agents built on the platform can now respond to webhooks over plain HTTP, not just Slack, and Slack conversations can carry file attachments such as contracts or spreadsheets directly into an agent’s context. Derek Gilbert, an engineer at Consensus, credited the platform’s Slack integration and its support for custom MCP servers, the protocol Anthropic created for connecting AI models to outside tools, with letting his team run “always on triage with access to the right monitoring and the ability to open PRs and alert teams.”
The built in search tool runs on Parallel, a web search provider, and LangChain is making it free to use while Managed Deep Agents remains in beta. Developers add it with a few lines of configuration instead of setting up a separate account and API key, and every search call shows up in LangSmith’s tracing dashboard alongside its cost and latency.
None of this makes LangChain the only company chasing the plumbing layer for enterprise agents. OpenAI, Anthropic and Microsoft are each building their own versions of the same authentication, memory and channel infrastructure into their agent tooling, and the winner of that race will likely be decided by which platform teams trust with live customer data first. For any team currently building a production agent in house, the relevant comparison is not whether Managed Deep Agents works, but whether the months saved on infrastructure outweigh the cost of depending on LangChain’s hosted layer once the free search and beta pricing end.
Reported by LangChain on 24 September 2026.
