Skip to main content
Agentic Enterprise.

Target-state architecture for the agentic enterprise

Seven planes across fourteen layers.

The components, protocols, control points, and contested choices, with the evidence behind each and the refusals stated plainly. Built for the architects who will design it, and written so that everyone who will live with it can follow: every page opens in plain terms, and every technical word explains itself.

The target-state architectureGo layer by layerWhat the experiments showed
  • Every claim dated and sourced
  • Vendor figures labelled
  • Negative results published
  1. 01

    Operate

    Execution plane

    Runtimes and durable sessions

  2. 02

    Operate

    Action plane

    Gateways and governed APIs

  3. 03

    Operate

    Knowledge plane

    Curated corpora and memory

  4. 04

    Govern

    Control plane

    Identity, policy, approvals, budgets

  5. 05

    Learn

    Improvement plane

    Evals, promotion, demotion

  6. 06

    Prove

    Evidence plane

    Independent traces and records

  7. 07

    Accountability

    Human plane

    Intent, accountability, exceptions

Enforcement and evidence must remain outside the agent’s influence.

The cross-layer spine

Six things every design must get right.

Each one cuts across every layer of the estate. Get these right and the layer-by-layer choices become ordinary engineering.

The four deterministic zones

Four places where a model may advise but never decide: access, money, physical safety, and official records. What sits in each, and how to build the boundary.

The identity and delegation chain

How every agent action stays traceable to the person who asked for it, from the agent's own identity through to the business system it touches.

Enforcement outside the model

Why the rules that stop an agent live outside the model, in a gateway and a policy engine, and why AI filters alone are not enough.

The data-to-memory pipeline

How company information reaches an agent: prepared, indexed, permission-checked, cited back to its source, and deleted everywhere when it must be.

The learning flywheel

How agents get better over time without anyone losing control: every change is tested, approved, rolled out in stages, and can be pulled back.

The autonomy contract

How much an agent may do on its own, decided per task, with the controls each level requires and the checks that must pass before autonomy rises.
Layer by layer

What you already own, and what changes in each part.

Fourteen deep pages. Each one states the target state, the mechanisms, the contested choices with verdicts, the cross-cutting concerns, and where the evidence runs out.

R01Infrastructure and computeR02Data platformR03Integration fabricR04Systems of recordR05Line of business and OTR06Intelligence and learningR07Agent platformR08Productivity and collaborationR09Experience and channelsR10Security and identityR11Governance, risk and sovereigntyR12Observability and FinOpsR13Supervision and oversightR14Agent data engineering
Tested, not asserted

We ran the experiments. Here is what held.

Five rounds on public corpora with human relevance judgements, seeded and reproducible to the byte. The multi-vector result held, the purpose-view result held under one condition, and the idea the programme was built to prove did not survive. All of it is published, with the reversals.

Read the findingsThe recommended approach
How we caught our own errors

One vector per document

0.815 to 0.294

nDCG@10 as a document grows from one aspect to ten. The single-vector index collapses.

Purpose views, aspect-targeted queries

+0.188

Over matched chunks at ten aspects per document, with 25 percent fewer embeddings.

Purpose views, whole-document queries

−0.032

Below matched chunks on human-judged scientific abstracts. The boundary of the method.

Objective conditioning

0 of 5

Forms of telling the index the goal that survived testing, including real human-judged data.

How this is written

Verdicts, sources, and gaps stated as gaps.

Evidence status on every claim

Vendor-published figures are labelled. Author positions are labelled. Unverified claims are excluded and the exclusion is stated.

Verdicts, not surveys

Every contested choice ends in a decision with the discriminator that decides your case, and what would change it.

Gaps published as gaps

Six cross-cutting concerns have no complete answer anywhere. They are listed with their status rather than papered over.
Ready when you are

Pick your layer.
Walk away with the design.

Start with the architecture25 decisions with verdicts

The Agentic Enterprise Architecture Guide