88% of agent pilots die because teams can't tell when a run quietly went wrong. Sentinel watches every run, scores its health 0 → 1, and explains exactly what broke. Framework-agnostic. Zero dependencies. Free, locally, in under 5 minutes.
A run "succeeds," returns nothing useful, loops 40 times, hallucinates a fact, or burns $40 — and nobody notices until a customer does. Uptime and error rates are blind to it, because nothing errored. Sentinel sees what they can't.
Six heuristic detectors run free and locally in the SDK. Two model-graded judges add semantic checks on the hosted tier — same Trace in → findings out contract.
REPEATED_TOOL_CALLSame tool + args N times — the agent is stuck in a loop.
EMPTY_OUTPUT"Succeeded" but returned nothing — the classic silent failure.
COST_SPIKEA run that crosses an absolute or relative cost threshold.
RUNAWAY_STEPSThe run blew past its step ceiling and won't terminate.
LATENCY_HIGHWall-clock exceeded threshold — a run that times out on users.
SPAN_ERRORA tool or step raised an exception mid-chain.
HALLUCINATIONJudge: the final answer asserts facts no step could support.
GOAL_INCOMPLETEJudge: the agent stopped early or answered a different question.
One package, zero dependencies — Python or TypeScript. No account, no backend.
One import for LangChain, LlamaIndex, CrewAI, or your OpenAI/Anthropic client. It never crashes your app.
See failures locally for free, or set an API key to sync to your dashboard and gate CI on regressions.
For developers — local, offline, forever.
For teams shipping agents to real users.
For orgs with security + compliance needs.
Free locally, forever. Add a key when your team needs the dashboard and CI gate.