LangChain Adds Connections So Agents Can Use Each Caller’s Identity

The new Managed Deep Agents feature keeps credentials in a LangSmith workspace and lets developers choose between shared deployment access and per-user authority.

By 2 min read
LangChain Adds Connections So Agents Can Use Each Caller’s Identity
LangChain Adds Connections So Agents Can Use Each Caller’s Identity

Listen to this story

The audio brief

About 1:32
0:001:32
Read transcript
An AI agent can now file a GitHub issue under the identity of the person who asked for it, instead of using one shared bot account. LangChain’s new Connections feature for Managed Deep Agents adds workspace-managed credentials, so developers can choose whether an agent acts with shared deployment access or with each caller’s authority. The feature is available in Managed Deep Agents version zero point seven point zero and later. A named credential is stored in a LangSmith workspace and retrieved at runtime through `connections.get()`, rather than being embedded in application code. The important distinction is ownership: a credential can belong to the agent, making access common to every caller, or to the user, making authorization specific to the person making the request. Either option can use a static secret or an OAuth grant. That user-owned path is where the practical change shows up. In LangChain’s GitHub example, the agent creates an issue as the requesting user. And if two users have different repository permissions, the same deployment can return different search results. The managed runtime handles OAuth authorization, token retrieval, refresh-related steps, and interruptions for consent, without requiring custom callback routes, token storage, or consent screens. Shared accounts are still available, including deployment-level OAuth. The catalog covers GitHub and 22 other services, plus custom OAuth metadata and MCP connections. The key design choice is now explicit: should the agent act as one shared account, or carry each caller’s authority into the action?

Story brief

3 key points

LangChain’s Managed Deep Agents now support workspace-managed Connections, letting deployments use either shared credentials or OAuth grants tied to the individual caller. In version 0.7.0 and later, credentials are retrieved at runtime rather than embedded in application code. This enables actions such as creating GitHub issues under each requester’s identity and producing access-aware results. The managed runtime...

  1. 01

    Connections is available in Managed Deep Agents version 0.7.0 and later.

  2. 02

    Credentials can be agent-owned for shared access or user-owned for per-caller authorization.

  3. 03

    User-owned OAuth flows can pause execution for consent without custom callback routes or token storage.

A shared AI agent can now file a GitHub issue under the identity of the person who requested it, rather than a common bot account. LangChain has introduced Connections for Managed Deep Agents, a workspace-managed credential system for shared deployment access and per-caller credentials.

Available in Managed Deep Agents version 0.7.0 and later, Connections stores a named credential in a LangSmith workspace. Tools retrieve it at runtime through connections.get(), rather than embedding the credential in the application build.

The credential owner determines whose authority applies

Connections treats ownership and credential format as separate choices. A credential can belong to the agent or the user, and either can be a static secret or an OAuth grant. Agent-owned access is shared by every caller; user-owned access resolves for the individual making a request.

In LangChain’s GitHub example, an agent that creates an issue with a user-owned connection uses the requesting person’s identity. That also means two people can receive different search results from the same deployment when their repository access differs.

OAuth steps move into the managed runtime

For user-owned OAuth connections, Managed Deep Agents handles authorization, token retrieval, refresh-related flow and authorization interrupts. When a caller needs to grant access, the runtime can pause the run instead of requiring the project to build callback routes, token storage or consent screens.

The feature also supports MCP connections when a server supplies OAuth metadata and registers the client. LangChain uses Linear’s MCP server as an example of a route that avoids supplying a client ID, client secret or app registration.

Shared accounts remain available

Per-caller identity is optional. Agent-owned secrets can be rotated or revoked in the workspace without changing code or redeploying. Developers can also use a deployment-level OAuth grant so every caller acts as one shared account. The Connections catalog includes GitHub and 22 other services, alongside support for custom OAuth provider metadata.

Managed Deep Agents entered public beta in August with identity positioned as a way to establish trusted user context and access boundaries. Connections adds a credential layer to that earlier identity model.

Sources

  1. langchain.comManaged Deep Agents is now in public beta
  2. langchain.comConnections: managed credentials and per-caller identity for Managed Deep Agents

Loading discussion...