LangChain Managed Deep Agents 0.8: User Memory, Webhook Channels

LangChain has shipped version 0.8 of LangChain Managed Deep Agents, the latest release of its managed infrastructure for putting production AI agents into service. The update tackles four recurring pain points for teams running agents in production: memory, authentication, channels, and tool management. The release landed on September 24, 2026, with details published on the LangChain blog.

The headline addition is user-level memory. Agents built on LangChain Managed Deep Agents can now remember context that is specific to each authenticated caller, keeping personal context separate from the shared memory the whole workspace sees. Alongside it comes a new HTTP channel, so agents are no longer bound to Slack, plus file transfer in Slack and a built-in web search tool powered by Parallel.

Managed Deep Agents, often shortened to MDA, combines the open Deep Agents harness with managed infrastructure. Developers define agent behavior, skills, and tools in a code-first project, and the runtime supplies durable execution, sandboxes, tracing, and evals. With the 0.8 release, the platform is trying to close the gap between an agent that works in a demo and one a business can actually operate.

Two memory layers: shared agent memory meets per-user memory

Previously, LangChain Managed Deep Agents offered one durable memory store for shared team knowledge. That worked for workspace-wide context, but it could not remember anything about individual users — or worse, it risked leaking one person's context into another's conversation.

Version 0.8 adds a second layer for user memory, keyed to the authenticated caller who starts each run. The runtime does not copy content between the two layers, and teams can set access policies at each level so user-specific information stays out of group conversations. By default, direct messages in Slack surface both memory layers, while group channels and HTTP requests only see shared agent memory. Teams can override those defaults.

Customer-facing agents were the motivating use case. According to the release notes, Zahid, chief technology officer at Kyth.ai, pointed to client confidentiality as the reason user-level memory matters: an agent serving many clients needs to keep each one's context strictly apart. That requirement has kept many teams from putting agents in front of customers at all.

The release also reworks how agents authenticate to external services. MDA 0.8 supports both agent-owned credentials and user-owned credentials for connections such as GitHub and Notion. Agent-owned credentials are shared across users. User-owned credentials let each person bring their own permissions, which matters when team members have different access levels on the same service. LangSmith manages authorization for 23 services, including Linear, GitHub, and Google Workspace, per coverage of the release.

Beyond Slack: HTTP channels and built-in web search

Until now, Slack was effectively the only way to talk to a Managed Deep Agent. The 0.8 release adds an HTTP channel: any service that can send a JSON webhook can start an agent run and receive the response. Internal portals, customer-facing support systems, order management tools — anything that can POST a JSON payload can now reach an agent. No Slack workspace license is required.

Slack itself gets richer too. File transfer support means users can send logs, spreadsheets, contracts, screenshots, and other documents directly to an agent inside a conversation. Derek Gilbert, an engineer at Consensus quoted in the release coverage, described the always-on triage potential for support teams.

Web search, one of the most common agent tools, is now built in. Managed Deep Agents 0.8 ships a prebuilt search tool powered by Parallel, so teams no longer need to set up a separate vendor account or API key. Search calls, latency, and errors appear in LangSmith traces alongside the rest of an agent's run, which makes it possible to debug a bad answer by tracing it back to the search that produced it. Parallel says its search runs with zero data retention and is free while Managed Deep Agents stays in beta.

Why it matters for the agent stack

The release lands in a week when every part of the agent stack is getting attention. Enterprise teams are counting their agents, putting runtime guardrails around them, and figuring out how agents should authenticate — the same week saw Nvidia open its agent safety platform around runtime policy enforcement, and Google push open-source orchestration for agent workflows.

LangChain is attacking a different layer of the same problem: the plumbing that turns a prototype into something operations can run. Identity-scoped auth, memory that respects who is talking, channels that meet users where they already work — these are the unglamorous features that decide whether an agent survives contact with a real enterprise.

Managed Deep Agents remains in public beta, LangChain says, and access requires a LangSmith Plus or Startup plan, according to coverage of the announcement. The built-in search is free only for the duration of the beta, the company says. The broader signal is that the agent infrastructure market is competing on the unglamorous parts — memory scoping, credentials, auditability — as much as on model quality.

Sources: LangChain blog, Parallel blog, Pondero, BadSignal.