Skip to content

deepagents@1.17.0

Choose a tag to compare

@github-actions github-actions released this 17 Nov 00:29
· 4 commits to main since this release
67d83a2

Minor Changes

  • #1058 06fa9a9 Thanks @byhow! - createDeepAgent stores messages as per-step deltas with a full copy every 50 updates, as Python deepagents does, instead of copying the whole conversation into every checkpoint. A 100-step run on MemorySaver stores 4.0 MB instead of 112.4 MB, and runs 5.5x faster per step. getState returns the same messages, and existing threads continue in place. A custom stateSchema keeps the delta channel unless it declares messages itself. Subagents defined by a spec still store full copies, while a deep agent passed as a subagent's runnable stores deltas.

    • Postgres users should upgrade @langchain/langgraph-checkpoint-postgres to 1.1.3 or later, which reads the deltas in a few paged queries.
    • Removing a message whose id is not in the state is now ignored instead of throwing.
    • The @langchain/langgraph-checkpoint floor rises to ^1.2.0. With SQLite, use @langchain/langgraph-checkpoint-sqlite 1.1.0 or later so parallel tool results reload in call order.
    • Downgrading deepagents after it has saved threads is safe only while @langchain/langgraph stays at 1.5.1 or later. Earlier versions read those threads as having no messages.

Patch Changes

  • #1075 eed9717 Thanks @eliornl! - fix(deepagents): stream async subagent runs to apps that join them

    The runs that the async subagent tools start on the same deployment, and the run that wakes the parent when a task finishes, now stream like an app's own runs. An app that joins one of these runs sees the reply token by token, including from subgraphs, and can replay events it missed. Subagents with a url keep their server's defaults.