Docs › LangGraph scheduled agent monitoring with RunVouch
LangGraph scheduled agent monitoring with RunVouch
LangGraph scheduled agent monitoring for the cycle that never converges: LangGraph gives you cycles by design. A cycle that never converges, on a nightly schedule, is a bill. LangSmith shows you the trace afterwards; RunVouch stops the run when it crosses the cap.
How it runs on LangGraph
Around the graph invocation with the Python client; report each tool node call with run.tool() so identical inputs are detected as a retry storm.
Store the key
An environment variable wherever the graph runs (cron, container, Lambda).
Wrap the job
import runvouch
from graph import app # your compiled StateGraph
runvouch.agent("research-digest", cadence_s=86400, cap_run_cost=2, evidence_required=True)
with runvouch.vouch("research-digest", evidence=lambda: {"digest": len(result["digest"]) > 200}) as run:
result = None
for event in app.stream({"topic": "..."}, stream_mode="updates"):
for node, update in event.items():
run.tool(node, update.get("tool_input", {}), cost=update.get("cost", 0))
result = app.get_state(config).valuesrv fails open: if RunVouch is unreachable the job still runs and you get one warning line.
Register the cadence and caps
rv agent nightly-report --cadence 24h --grace 30m --max-runtime 1h --evidence --cap-run-cost 2
Register the agent once, from anywhere with the key. Cadence is what turns a schedule that stopped into an alert; --evidence makes a run without evidence a failure; the caps pause the agent when it overspends.
What goes silent on LangGraph
- Cycles with a high recursion limit look like progress; 40 identical tool inputs in one run is RETRY_STORM.
- A graph that ends with an empty final state is a completed run; evidence says whether the digest exists.
- Nothing in LangGraph knows the schedule; MISSED comes from the agent's cadence.
What LangGraph does not tell you
LangGraph raises GraphRecursionError when the recursion limit (default 25 steps) is reached; anything below that limit is a normal run, and the limit is often raised in exactly the graphs that loop. LangSmith tracing records every step for inspection after the fact and can alert on error rate and latency; it does not know the graph was supposed to run tonight, and it does not stop a run mid-way.
What RunVouch detects
| Alert | What it means here |
|---|---|
| MISSED | cadence plus grace passed and no run started |
| FAILED | non-zero exit or a reported failure, with the stderr excerpt |
| NO_EVIDENCE | the run said ok but the file, URL or assertion you required is missing |
| STALLED | a run started and never ended within max runtime |
| RETRY_STORM | the same tool called with identical input many times in one run |
| BUDGET_RUN / BUDGET_DAY | cost cap crossed; the agent is paused until you resume it |
| DRIFT | duration or output size far off its 7-run baseline |
Set up in two minutes
- Get a free key (3 agents, no card) and store it where this page says.
- wrap the job: copy the snippet above into the scheduled job.
- Register the cadence once:
rv agent nightly-report --cadence 24h --grace 30m --evidence, or let the first run create the agent and set the cadence on the dashboard. - Send one test alert:
curl -X POST https://api.runvouch.com/v1/settings/test-alert -H "X-API-Key: $RUNVOUCH_KEY". The next missed, failed or empty run reaches the same channels.
Also running agent frameworks?
CrewAI. AutoGen. OpenAI Agents SDK. Pydantic AI and smolagents. The full list of runtimes is on integrations, and the same agent can report from several of them: one job, one cadence, wherever it happens to run.
Need a key? Get a free key · Stuck? contact