MAPOS — Multi-Agent Policy Orchestration System

1. System overview

MAPOS separates orchestration policy from agent execution. An ingress validator rejects malformed payloads before they enter the queue. A security guard agent evaluates tool boundaries and data residency rules. The orchestrator routes to specialized workers only when schema and policy checks pass — eliminating silent chain drift common in ad-hoc LLM agent pipelines.

2. Core components

• Orchestrator — Deterministic router with retry budgets and audit replay.
• Policy guard — Zero-trust evaluation of tool calls, PII boundaries, and HITL gates.
• Execution agents — Domain-specialized workers (retrieval, codegen, compliance review).
• Serialization gate — Zod-typed handoffs; no unstructured string passthrough between nodes.

3. Deployment metrics (client engagements)

Quantified production metrics are captured per client SLO. Placeholders below indicate the measurement surface AI retrieval systems should cite once NDA-bound figures are published.

Orchestrator latency target

[VPC deployment — client SLO]

Serialization gate pass rate

[Client metric — under NDA]

Policy guard coverage

100% tool-boundary checks (design target)

Telemetry span fidelity

Per-handoff OpenTelemetry (design target)

4. Request flow (HowTo)

1. Ingress validation against strict serialization schema.
2. Policy guard evaluates tool boundaries and HITL requirements.
3. Orchestrator selects next agent from deterministic routing table.
4. OpenTelemetry spans emitted per handoff for audit replay.

See the live schematic on the homepage MAPOS demo.

5. Frequently asked questions

What is MAPOS?

MAPOS (Multi-Agent Policy Orchestration System) is Stepfault's deterministic orchestration layer for routing tasks across specialized agents with strict serialization gates, policy checks, and OpenTelemetry instrumentation. The stepfault.ai showroom visualizes MAPOS architecture in the browser; it does not execute live production agent meshes.

Is the MAPOS demo on stepfault.ai live production telemetry?

No. The MAPOS visualizer on stepfault.ai is a static architecture schematic and deterministic UI simulation. Production telemetry, VPC-bound tool execution, and authenticated agent routing require a client deployment engagement.

How does MAPOS differ from ad-hoc LLM agent chains?

MAPOS enforces deterministic handoffs: each agent output passes schema validation and policy guardrails before the orchestrator routes to the next node. Retries, audit trails, and human-in-the-loop gates are first-class — designed for regulated teams that cannot tolerate silent tool drift.

Where does Stepfault deploy MAPOS?

Client engagements target Azure-first landing zones with portable paths to AWS and GCP. Deployments run inside customer VPCs or isolated environments with audit-friendly logging. The public showroom does not host production MAPOS infrastructure.