WORKFLOW RELIABILITY, WITH A CONTROL PLANE

Diagnose the signal.
Contain the noise.

Six bounded agents turn n8n, Make, Zapier, and custom execution evidence into a deterministic incident state, a normalized fingerprint, and an approval-gated runbook.

Open diagnostic ↘ Inspect OpenAPI Try with your data

6: bounded agents
18: evaluation measures
17: trace dimensions
0: automatic fixes

RELIABILITY SIGNALLIVE REPLAY

INGESTverified

FINGERPRINTcredential_or_auth

POLICYapproval required

trace_id · evidence_hash · graph_version

01 / FAILURE

One broken workflow can hide behind a green dashboard.

Success status, stale heartbeats, duplicate webhooks, and zero-item runs can all mislead an operations team.

02 / DECISION

Business truth is deterministic.

Rules classify coverage, freshness, retries, maintenance, and recovery. An optional model only rewrites verified output.

03 / BOUNDARY

The public system never touches a workflow.

Every operational runbook is dry-run. Approval records intent. A private deployment must bind an audited connector.

LIVE INCIDENT LAB

Change the evidence.
Change the diagnosis.

These presets hit the same typed FastAPI route used by Postman. Every result is produced by the LangGraph council, not swapped front-end copy.

Execution packet

Typed and source-labelled

PYDANTIC VALIDATED

Workflow reference Platform

Execution status Source observed

Minutes since success Expected interval

Latency, ms Baseline, ms

Error message

Additional reliability evidence

Execution ID Failed node

Error type Items processed

Retry count Previous health

Duplicate seen Maintenance window

Untrusted notes

Replay mode costs $0 and cannot call or modify an external platform.

Council trace

Waiting for evidence

READY

◇

Run a scenario to inspect each agent decision and evidence-backed transition.

CONTROLLED MULTI-AGENT RUNTIME

Parallel diagnosis.
One governed decision.

LangGraph makes every branch, checkpoint, and pause explicit. The model cannot redefine severity, evidence, or tool permission.

BOUNDARYFastAPIschema · auth · tenancy · idempotency

→

PARALLEL3 diagnostic agentsingestion · fingerprint · heartbeat

→

COORDINATEIncident statepriority · evidence · runbook

→

CONTROLInterrupt gateapprove · reject · audit

Why REST

Incident ingestion and decisions are command-shaped. REST keeps the contract browser and Postman friendly without GraphQL query complexity.

Why deterministic fingerprints

Credentials, timeouts, rate limits, and schema drift must group the same way on every replay. An LLM may narrate, but cannot classify.

Why LangGraph

Parallel branches, bounded steps, persisted state, and an interrupt-based approval are visible and testable instead of hidden in free-form agent chat.

Why not automatic repair

A retry can duplicate side effects and a credential change can break every consumer. The public system deliberately stops at a dry-run runbook.

EVALUATION AND OBSERVABILITY

Prove behavior before
trusting a runbook.

Golden scenarios measure classification, control enforcement, evidence grounding, recovery, and resilience. Traces cover the whole run without capturing raw content by default.

EVALUATION PROTOCOLloading

Reading the published golden report.

OBSERVABILITY PROTOCOL17 dimensions

LangSmith graph traces, OpenTelemetry spans, structured logs, and optional Sentry errors share one run ID.

 run → ingest ┬ fingerprint
             ├ heartbeat
             └ policy interrupt
content.capture = false

DEPLOYMENT BOUNDARY$0 demo

One Docker Compose command
Vercel synthetic walkthrough
PostgreSQL production checkpoints
Optional OpenRouter narration
No paid fallback by default