Migrating from LangGraph

Migrating from LangGraph to Lumen + Kova

SqliteSaver → LumenCheckpointer + LumenTracer, deploy your Agent to a Kova cluster—no changes to your business code.

What you get

Before (plain LangGraph)	After
SqliteSaver single-node persistence	LumenCheckpointer + Kova microsecond-level WAL recovery
No tracing	LumenTracer automatic capture
No cost tracking	CostTracker aggregated per node
Manual replay	Replay.from_run_id() in one line

Migration steps

1. Replace the Checkpointer — LumenCheckpointer implements BaseCheckpointSaver 100%, with no changes to your business code.

   - from langgraph.checkpoint.sqlite import SqliteSaver
   + from lumen_ai import LumenCheckpointer
   
   - graph = wf.compile(
   -     checkpointer=SqliteSaver.from_conn_string("./state.db")
   - )
   + graph = wf.compile(
   +     checkpointer=LumenCheckpointer()
   + )

2. Add the Tracer — once running locally, visit http://localhost:7070 to see the trace timeline.

   + from lumen_ai import LumenTracer
   
     graph = wf.compile(
         checkpointer=LumenCheckpointer(),
   +     callbacks=[LumenTracer()],
     )

3. Deploy to Kova

   # agent.yaml
   name: research-agent
   entry: python my_agent.py
   checkpointer: lumen
   replicas: 3

   lumen deploy --target kova://prod-cluster

Local verification

python my_agent.py           # runs normally
kill -9 $(pgrep -f my_agent) # simulate a crash
python my_agent.py           # resumes from the interruption point, no LLM re-call

Performance comparison

Metric	Before	After
Recovery latency	8ms	microsecond-level
Multi-process	needs locking	native
Cross-datacenter	manual	async replication
LLM re-call	sometimes	never

Rollback

Keep the old SqliteSaver code commented out; revert the diff at any time to return to the original state. The data is not shared, but they can run in parallel.

Next steps

Next steps

Lumen Python SDKDeploy KovaFull tutorial