Skip to content

What changed / migrating

Two changes landed in June and July 2026: the family was renamed to LangStage, and the shared core adopted the AG-UI protocol as the one way every surface streams a turn. Both are designed to be a non-event for existing users — old installs, imports, and commands keep working through a transition window.

What can break

  • Python ≥ 3.11. The shared core requires 3.11 (it was 3.10), because AG-UI's streaming stack doesn't work on 3.10. The surface packages already required 3.11. On 3.10, stay on the pre-1.0 langgraph-stream-parser 0.3.x or upgrade Python.
  • Direct use of the old event layer. If your own code imported StreamParser, event_to_dict or stream_graph_updates, it needs the small migration below.

Everything else (package names, imports, commands, config names) is shimmed or falls back.

Name map

Old package New package
cowork-dash langstage
deepagent-code langstage-cli
deepagent-lab langstage-jupyter
deepagent-vscode langstage-vscode
deepagent-hermes langstage-hermes
langgraph-stream-parser langstage-core (the shared core; the old name still installs as a shim)

All are on PyPI — pip install <new package> gets the current release. (This table intentionally lists no version numbers; a hand-maintained "latest" column only goes stale — check PyPI for the current version of each.)

What still works

  • pip install <old-name> — each old name has a final "tombstone" release that simply depends on the new package, so installing it pulls the rename.
  • import <old_module> — e.g. import cowork_dash, import deepagent_code, and the documented host spec deepagent_hermes.agent:graph still import, via a deprecated alias package that emits a DeprecationWarning.
  • Old console commands — cowork-dash, deepagent-code, deepagent-lab, deepagent-hermes remain as aliases of the new commands. (deepagent-hermes prints a one-line deprecation note: on each run.)

These shims are kept for a transition window, then removed in a future major. Update when convenient — nothing breaks today.

Config vocabulary

The canonical config vocabulary is now LANGSTAGE_*:

Old New
DEEPAGENT_* env vars LANGSTAGE_*
deepagents.toml langstage.toml
~/.deepagents/config.toml ~/.langstage/config.toml
DEEPAGENTS_CONFIG_HOME LANGSTAGE_CONFIG_HOME

The legacy names still resolve as a deprecated fallback. The canonical name wins when both are set, and each legacy name prints one note: line on stderr per process (LANGSTAGE_SUPPRESS_LEGACY_NOTICE=1 silences it). Moving the global config out of ~/.deepagents/ also avoids a collision with LangChain's dcode, which now owns that directory.

langstage-hermes keeps its existing home dir for safety: an existing ~/.deepagent-hermes directory keeps being used if present, so no skills, memories, or session history are orphaned by upgrading. The legacy deepagent-hermes.toml file name is still read, with a deprecation note.

Every surface now streams over AG-UI

The shared core adopted the AG-UI protocol, the event-based wire for streaming text, tool calls, reasoning, state and human-in-the-loop to frontends. It started as an opt-in and is now the only streaming path: every surface drives your graph through the core's in-process AG-UI bridge, and langstage-core[agui] is a regular dependency of each one. You don't need to change your agent. What you get:

  • The same frames on every surface, with the same rules. See the frame reference.

  • langstage-agui, which serves any CompiledGraph as an AG-UI HTTP endpoint:

    pip install "langstage-core[agui]"
langstage-agui --agent my_agent.py:graph     # or --demo

    See Serve over AG-UI.

Notes for AG-UI client authors:

  • Graphs compiled without a checkpointer work: one is attached for you (AG-UI needs threaded state for interrupts and resume).
  • Agent failures surface as a terminal RUN_ERROR event rather than a dropped stream.
  • Use a new message id per turn. The adapter dedupes messages by id, so a reused id silently drops the later turn.
  • Interrupts arrive as a CUSTOM on_interrupt event. Resume by starting a new run on the same thread with AG-UI's standard RunAgentInput.resume (ag-ui-langgraph 0.0.43 and later). The older forwardedProps.command.resume still works as a fallback but is deprecated. The decision payload is described in Human-in-the-loop.

If you used langgraph-stream-parser directly

langstage-core 1.0 is the rename of langgraph-stream-parser. import langgraph_stream_parser keeps working only while the separate langgraph-stream-parser compat package is installed (upgrading it in place keeps it; a fresh pip install langstage-core doesn't pull it). New code should import langstage_core.

The old event layer was removed in 1.0. If you used it, migrate:

Removed Use instead
StreamParser, langstage_core.events, event_to_dict langstage_core.agui.iter_event_frames / iter_chunk_frames (frame dicts, same vocabulary)
stream_graph_updates, resume_graph_from_interrupt iter_chunk_frames(agent, message, thread_id, resume=...)
adapters.CLIAdapter, PrintAdapter, FastAPIAdapter, JupyterDisplay SessionAdapter (in-process) or build_app / serve (HTTP)
from langstage_cli import stream_graph_updates, prepare_agent_input the iter_chunk_frames snippet on the Terminal page

Kept and unchanged: load_agent_spec, HostConfig, prepare_agent_input, create_resume_input, the task engine and the extractors. The rationale is in ADR 0003.

  1. pip install langstage langstage-cli langstage-jupyter langstage-vscode langstage-hermes (only the ones you use).
  2. Swap import paths and commands to the new names at your leisure.
  3. Rename deepagents.toml → langstage.toml and DEEPAGENT_* → LANGSTAGE_* when convenient; run any stage's --show-config to confirm what's resolving.