Jump to another pillar

Pillar 02 of five

Agentic product and multi-agent architecture

What you ship

  • What the customer opens: the work queue, the studio, the library, the chat
  • What their own agents may read through MCP, and what only your controlled write path may change
  • Fine-tuned open-weight models based on aggregated and anonymised datasets. Released under a licence TBD

What runs underneath

  • The domain graph, and the arithmetic no model is allowed to touch
  • The agent roster, the families of models required and the evaluation framework
  • Fusion gateway: one path to every model (open-weight or closed), so no agent names a provider

VAO · pillar 02 · what runs underneath

The stack, by layer#

Nine layers an orchestration vendor has to choose deliberately.

Each tool carries a maturity word — how far this choice has been taken, not how well-sourced a claim is. It is a separate scale from the [E] / [P] / [I] evidence tags used elsewhere on the site.

— running code and traces
— compared against alternatives, not yet built
— the choice I would make, untested

What decides

01

Orchestration

Holds the graph of steps, the state between them, and the pause-and-resume path when a human has to approve something.

LangGraph
02

Fusion gateway

One path to every model, so no agent names a provider. Routing by task class and cost, and one model family checking another's work before anything binding.

LiteLLMvLLM
03

Models

Three classes, from frontier reasoning down to small and fast, plus the open-weight class that makes sovereign deployment possible.

FrontierMid-tierSmall / fastOpen-weight
04

Specialisation

Where taught judgement is compiled into a model rather than held in prompts.

PlatformsTogether AIvLLM + LoRA
TechniquesSFTDPOGRPO

What it knows

05

Retrieval

Finding the right prior case in years of correspondence, in several languages, where half the meaning sits in part numbers and reference codes.

Multilingual embedderBM25 / SPLADECross-encoder rerank
06

Governed connector layer

Not just the protocol. Who may reach which system, whether this specific call is allowed right now, whether it was logged, and what re-qualifies a tool when its description changes.

ProtocolMCPA2A
GatewayMCPXDocker MCP GatewayIBM ContextForge
Registry · identity · scanningMCPJungleOAuth 2.1MCP-Scan
07

State and stores

The domain graph, the checkpointed run state, the retrieval indexes, and the versioned artefacts.

PostgresGraph storeVector indexArtefact registry

What proves it

08

Observability and evaluation

Traces of what each agent saw, chose and cost. Then three classes of test — taught-routine suites, whole-loop runs in a synthetic environment, and calibration.

LangfuseArize
09

Independent audit

A third party given the trajectories, the methodology and access to the team — which then verifies the score and reviews the hardest cases.

Quesma

Open weights, and the two sovereignty questions#

Ask where the data physically runs. Then ask, separately, who could be legally compelled to hand it over. Two questions, two different answers — and conflating them is the most common error in a sovereignty pitch.

PlayerDeployment sovereigntyLegal sovereigntyWhat it is the right answer to
MistralFR · Apache 2.0clearsclearsThe default for a French or EU-jurisdiction mandate. On-premise, air-gapped or EU trusted-cloud, and the only one of the four with live defence-sector deployment.
Cohere / Aleph AlphaCA · DEclearsclearsRegulated-sector accounts wanting a production track record rather than a model release. Check which licence variant a proposal names — the earlier one is non-commercial, and the terms are not interchangeable.
NVIDIA NemotronUS · open weightsclearsdoes notLong-horizon agentic work — the class this architecture actually stresses. Best-documented training data of the four, which strengthens a never-trained-on-your-data commitment when a compliance team asks what the base model itself saw.
IBM GraniteUS · Apache 2.0clearsdoes notAudit-trail and indemnification-first accounts where vendor nationality is not the binding constraint. IP indemnification is a standard term, which none of the other three offer by default.

Where it runs#

Sovereignty is only real if the deployment topology supports it. Four of them — and the last is the one to design for first even though it sells last, because everything that makes it possible relaxes into the other three. The reverse is a rewrite.

AVendor cloudMulti-tenant, EU region, hosted model endpoints. Pilots and mid-market.
BCustomer cloudSingle tenant in the customer's own cloud account. Most enterprise accounts.
COn-premiseCustomer data centre, self-hosted open weights, tenant-local traces, no export.
DAir-gappedNo egress at all. Offline evaluation, mirrored registry, local calibration monitoring.