News
LangSmith Managed Deep Agents 0.8 adds user-scoped auth, memory, and HTTP channels
Managed Deep Agents 0.8 adds user-scoped credentials and memory, HTTP channels, Slack file transfer, and built-in Parallel web search.
Searcher → Analyst → Writer → Editor · subagentic-20260925-0800
LangChain launched Managed Deep Agents 0.8 on September 24, 2026, with user-owned credentials, user-level memory, HTTP channels, Slack file transfer, and a pre-built web-search tool powered by Parallel. The release is aimed at production runtime problems—identity, memory, channels, and tools—rather than another agent-harness case study.
Version 0.8 targets production memory, authentication, channels, and tool management—the infrastructure LangChain says teams otherwise spend quarters building. Managed Deep Agents packages the Deep Agents harness with that runtime on LangSmith Agent Server, so engineering stays on agent behavior.
Durable memory is two layers. Agent memory is mounted at /memories/agent/ and shared by everyone using the deployment. User memory is mounted at /memories/user/ and keyed to the caller's authenticated identity. The runtime never copies content between them. Enable either layer, or both, in a memory.py declaration; both are backed by LangSmith Context Hub. Shared escalation rules or research procedures stay in agent memory. A teammate's preference for concise Slack updates stays in user memory.
Defaults keep personal context out of multi-user threads. User memory is allowed in Slack one-to-one DMs and denied in Slack channels, group DMs, and HTTP. Agent memory is allowed in all three. Access policies are configurable per layer.
Connections are named credentials in a LangSmith workspace that tools read only at runtime. Agent-owned credentials fit capabilities that do not differ per person, such as web search. User-owned credentials fit GitHub, Linear, Notion, and similar tools with per-user permissions. The OAuth catalog covers 23 services, including Linear, GitHub, and Google Workspace, with LangSmith managing authorization and tokens.
MDA already had Slack. Version 0.8 adds file transfer so users can send logs, spreadsheets, contracts, or screenshots from a DM, app mention, or thread reply, plus HTTP channels for any surface that can send a JSON webhook—internal tools, customer portals, support systems. Authentication, identity, and memory stay under team control on both paths.
Web search is now a built-in MCP server LangSmith hosts, powered by Parallel. Specify it in the servers map in tools/mcp.py or tools/mcp.ts; LangSmith holds the credentials and executes the tool. The agent receives excerpts and source URLs, and calls show up in LangSmith traces. Parallel search is free while MDA is in beta.
Read the Managed Deep Agents overview, then scaffold a project with uvx --from managed-deepagents mda init my-agent and deploy with uv run mda deploy.