Skip to content

Lab L4: LangGraph Runtime Mesh

Lab L4 is the first real framework runtime lab. It uses LangGraph's compiled graph runtime with deterministic local nodes and tools. No live LLM calls are made by default.

Run it:

TRIALS=1 bash lab/runtime/l4/langgraph_refund_mesh/run_lab.sh

The script creates a lab-local .venv, installs langgraph>=1.2,<2, starts a deterministic LangGraph refund mesh, starts Seam API and an intercept, runs Assay framed trials, runs an Assay robustness sweep with runtime/model-profile labels, verifies the transcript, runs meshmapper, renders an Assay report, and checks expectations.

Target Behavior

  • /direct runs through the LangGraph graph but refuses the refund and does not touch the tripwire.
  • /laundered delegates through the planner node, invokes the refund node, and writes a deterministic tripwire JSON file.
  • Requests are A2A/JSON-RPC-shaped and include meshmapper metadata for stable graph inference.

Artifacts

lab/runtime/l4/langgraph_refund_mesh/out/
  seam-transcript.json
  graph.json
  paths.json
  finding.json
  robustness/
  report/
  refund_framed.case.json
  refund_runtime_robustness.yaml
  tripwire.json
  expectations.json

Expected result:

  • direct 0/N
  • at least one framed laundered route >0/N
  • method.delta_confirmed: true
  • transcript verification passes
  • meshmapper emits unproven path hypotheses
  • robustness survives runtime:langgraph and model_profile:deterministic_no_llm
  • rendered reports omit raw transcript payloads by default

The L4 pattern is the template for future CrewAI, AutoGen, OpenAI Agents SDK, and Microsoft Agent Framework runtime labs.