AWS Machine Learning vs LangGraph

The feed carries the LangGraph monorepo's per-package release tags — the core library, the SDK, the CLI, and three checkpoint backends — each publishing a raw commit list under a version-only title. Nearly all recent movement sits in checkpoint persistence: delta-channel history correctness, namespace matching scoped to segment boundaries, and an opt-in flag to skip expired rows on read. The newest tag is the Python SDK, whose own contribution is a decrypt replacement result and the ability to clear a cron end_time by passing None; the rest of its body is the monorepo's release history restated. The core library's own changes are dependency bumps plus a tracing API exposed, stripped of tags, deleted, and exposed again across three releases.

Checkpointing — how agent state is persisted and replayed — is where the engineering attention concentrates, and the specific fixes are the kind that only surface once people run long-lived graphs against real databases rather than in notebooks. The SDK's cron and decryption work points at the same population: scheduled, long-running deployments with encrypted state. The second thread is that TracePolicy has not settled, which puts the observability surface visibly still in design. Neither thread changes what LangGraph is for; both are the work of making a 1.x framework survive production use.

The checkpoint packages will keep releasing in lockstep with the core library, since a single change routinely fans out across three tags. Whether trace_policy survives this time is the open question these entries do not answer.