LangGraph 1.2 introduces DeltaChannel, a new primitive that solves the O(N²) checkpoint storage problem for long-running agents. Instead of serializing full state snapshots at every step, DeltaChannel stores only the diff per step and writes full snapshots every K steps (default: 50). For a coding agent running 200 turns, this reduces checkpoint storage from 5.3 GB to 129 MB — a 41x reduction. Deep Agents v0.6 ships with delta-backed messages and files by default, requiring no configuration changes or data migration. The API adds a DeltaChannel type with a reducer function and snapshot_frequency parameter, with a stricter batching-invariance contract on reducers. All existing LangGraph features (interrupts, time-travel, tooling) remain unchanged.
