The autonomous agent platform built for production

Contracts are validated at startup and enforced by CI, so breaking changes are caught before they reach production. The schema covers input types, output shapes, FSM transitions, and allowed side-effect categories — giving you a machine-readable spec for every unit of work in your pipeline.

Each orchestrator owns a FSM contract that lists every valid state, event, and transition guard. Illegal transitions are rejected at the framework level — not caught by ad-hoc if-else chains in your business logic. This makes workflows reproducible, testable, and safe to parallelize.

The boundary is not a convention — it is a Kafka topic ACL and a CI check that fails the build if any Effect node issues a write without a valid service token. Audit logs are emitted for every mutation so your compliance team always has a full trail.

Events carry structured payloads: node ID, correlation ID, tenant, duration, input hash, and output schema fingerprint. Omnidash aggregates them into a live topology view so you can pinpoint bottlenecks, replay failed runs, and set latency alerts — all without touching your node code.

Tenant isolation is enforced via PostgreSQL RLS policies seeded directly from JWT claims. A compromised API key cannot cross tenant boundaries. Usage is tracked per-tenant with soft limits and overage alerts, giving you the primitives to build a metered SaaS product.

Every request carries a JWT validated by the platform's Keycloak realm. Tenant isolation is enforced at the database row level via RLS policies seeded from the JWT claims. API keys are scoped to a service account and a permission set, so a compromised key never grants cross-tenant access.

Patterns follow a lifecycle from candidate to provisional to validated. The engine clusters similar patterns, deduplicates them, and retires deprecated ones — so the context Claude receives reflects your actual current conventions, not last quarter's ones.

Memories are stored in Qdrant with semantic embeddings for similarity retrieval. A relationship graph in Memgraph tracks how intents, patterns, and outcomes relate to each other. Memory lifecycle management handles consolidation, deduplication, and archival automatically.

Omnidash subscribes directly to the Kafka event bus. Every node execution, routing decision, and pattern discovery event flows through without additional SDK calls. Dashboards include P50/P95/P99 latencies, tenant usage breakdowns, correlation traces, and live topology graphs.

Redaction runs at the event emission layer before any data touches Kafka. Patterns cover OpenAI keys (sk-*), AWS keys (AKIA*), GitHub tokens (ghp_*), Slack tokens (xox*), PEM certificates, Bearer tokens, and passwords in URLs — with a dev-mode log of redacted fields for auditability.

The event bus routes failed messages to a dedicated DLQ topic. Each entry carries the original payload, the error, the node version, and the correlation ID. Retry policies are defined in the node contract — not hardcoded — so you can tune per-node without touching orchestration logic.

The handshake system fingerprints event schemas, Kafka topic contracts, and API surface areas at every PR. A mismatch between producer and consumer contracts fails the build in the consumer repo, not in production. Schema migrations are blocked by a freeze flag until the dependent upgrade window is confirmed.