Four blocks, one trust spine

Orchevo separates the platform into four blocks so that building, running, governing and deploying agents are independent concerns. Capability is created in the Agentic Factory, served by Agentic Services, governed by one trust spine that applies to every action, and carried anywhere by a Kubernetes foundation.

Block 01 · Build — the Agentic Factory

Business users assemble agents with a guided wizard or describe them in plain language; developers design visual DAG flows, work from IDEs and notebooks, or bring agents built on LangGraph, CrewAI or AutoGen through adapters. Whatever the path, the output is one agent manifest — model policy and fallbacks, tool grants (MCP scopes), knowledge scopes with ACL inheritance, guardrail bindings, an evaluation suite and red-team probes, resource limits and escalation policy.

An evaluation harness runs golden sets and red-team probes on every version; a policy-enforced approval gate reviews manifest diffs and tool grants; GitOps promotes dev → staging → production. Production agents are immutable, and rollback is a pointer change — what you evaluated is exactly what runs.

Block 02 · Run — Agentic Services

Agentic Services consume Factory agents and models to pursue goals: a supervisor decomposes and routes to specialists; deterministic DAGs handle known processes with human-approval nodes; event-driven agents subscribe to business events for ambient automation. Sandboxed execution, scoped identity, short-lived credentials, explicit egress and spend limits mean that whatever an injected prompt convinces a model to try, the runtime physically cannot exceed approved permissions.

No agent calls a model directly

Every inference flows through one AI gateway — routing by task class, cost, latency and data classification; fallback chains; semantic caching; PII and injection rails at the choke point; and full audited logging. Retrieval is zero-copy and permission-aware: embeddings and pointers move, documents don't, and the caller's identity is pushed into the index so agents surface only what their user could open.

Block 03 · Govern — the trust spine

A single policy engine (RBAC + ABAC) enforces at four points — Studio, orchestration, gateway and data. Approval gates guard promotion, tool grants and high-impact actions. Every prompt, retrieval, tool call and response is immutably audited in-perimeter. Explainability and agent cards map to the EU AI Act and NIST AI RMF, and evaluations run in production with drift alarms. See the full Trust & Security page for the control detail.

Block 04 · Deploy — the Kubernetes foundation

• On-premises — full stack in your data centre; operator-owned GPUs; highest control.
• Private VPC — dedicated deployment in your cloud tenancy.
• Hybrid dual-zone — sovereign zone on-prem; general zone bursts to cloud via the gateway.
• Air-gapped — no connectivity; signed offline bundles; local telemetry only.
• Sovereign AI cloud — telco/MSP-operated, multi-tenant, co-branded national service.

Operators and GitOps (KServe/vLLM, Argo CD) make every environment — including the air-gapped one — reproducible from Git. Promotion is a merge; rollback is a revert.