Agentic platform engineering, with the platform still in charge.

One request, eight acts, every verdict computed live.

Every policy verdict below came from running the real conftest binary against the real Rego in policy/. Every manifest came from the real score-k8s binary and the provisioner set in platform/. Run it twice and the numbers are identical — that is what a control plane is supposed to give you.

Intent in, auditable pull request out.

Two front doors onto one golden path: a portal for humans, MCP for agents, the same authorization policy behind both. Every box is a real open-source project or a file in this repository — nothing is a placeholder.

The policy gate, and the agent that closes it.

17 controls across three bundles — 34 deny/warn rule bodies — all real Rego. Every message carries a policy id and an explicit remediation after ->. That second half is what makes the loop agentic rather than merely a linter: the agent parses it, changes a Score parameter, and the whole artefact is regenerated.

╭──────────────────────────────────────────────────────────────────────────────────────────────────╮
│ AGENTIC PLATFORM ENGINEERING — NORTHWIND RETAIL                                                  │
│ one request · eight acts · every verdict from a real OPA evaluation                              │
╰──────────────────────────────────────────────────────────────────────────────────────────────────╯

━━━ ACT V  The gate: what the platform cannot decide for you ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

  reasoner: deterministic reasoner (no model, fully reproducible)
  engine:   conftest 0.69.0  ·  real Rego, real evaluation

  ▍ ITERATION 1   $960.82/mo vs $400.00 budget
   1. NW-PCI-003  SQLInstance/nw-payments-ledger-prod: backupRetentionDays=7 is below the 30-day
                  PCI floor
      fix → set spec.parameters.backupRetentionDays>=30
   2. NW-FIN-001  payments-ledger/prod: estimated $960.82/mo exceeds the CC-4471 budget of
                  $400.00/mo by $560.82
      fix → right-size the largest line item or request a budget increase from the cost-center
            owner
   3. NW-FIN-003  payments-ledger/prod: resource "payments-ledger-postgres" is $820.67/mo — 85% of
                  the whole service budget
      fix → single line items over 70% need an explicit exception annotation from the cost-center
            owner

    agent → 2 input change(s), then re-render from source
      · db_backup_retention_days 7 → 30
        PCI requires a 30-day minimum recovery window
      · db_instance_class db.r6g.xlarge → db.t4g.large
        largest reducible line item; db.r6g.xlarge -> db.t4g.large stays inside the cost-centre
        envelope

  ▍ ITERATION 2   $463.15/mo vs $400.00 budget
   1. NW-FIN-001  payments-ledger/prod: estimated $463.15/mo exceeds the CC-4471 budget of
                  $400.00/mo by $63.15
      fix → right-size the largest line item or request a budget increase from the cost-center
            owner

    agent → 1 input change(s), then re-render from source
      · db_instance_class db.t4g.large → db.t4g.medium
        largest reducible line item; db.t4g.large -> db.t4g.medium stays inside the cost-centre
        envelope

  ▍ ITERATION 3   $359.15/mo vs $400.00 budget
    ✔ 0 denials across 4 gates


    PASSED — 0 denials    converged in 3 iterations


  Note what the agent never did: it never edited a rendered manifest
  to make a check go green. Every fix changed an input and the whole
  artefact was regenerated. Patch the output instead and the finding
  comes back silently on the next render — that is how automated
  remediation becomes worse than none.

• It fixes causes, never symptoms. Every remediation edits a Score parameter and the artefact is regenerated. Patch the rendered output instead and the finding returns silently on the next render — that is how automated remediation becomes worse than none.
• Compliance outranks cost. The budget gate reduced capacity and left the 30-day PCI backup window alone. It went the other way, 7 days to 30, in the same run.
• It stops when it cannot map a denial to an input. No guessing, no silencing — the finding escalates to a human, which is the correct outcome rather than a bug.
• The engine is the real one. The same conftest you would run in your own CI, at a pinned version, over Rego you can read in policy/.

Build the platform one control at a time

Start from what the model handed back and switch on the guardrails a platform team would encode once. Each verdict list is a genuine conftest + kube-linter run captured at build time by src/build_playground.py — the page replays recorded output, it does not approximate Rego in JavaScript.

A prompt is a request. An absent tool is a fact.

Every tool this platform exposes declares a permission, and tools/list is filtered by the calling identity before the model ever sees it. An agent without delivery:approve is not instructed to avoid approving production. It is never told the capability exists.

identity: agent:platform-agent  (12/14 tools visible)

    catalog.get_entity                 catalog:read
    catalog.query                      catalog:read
    catalog.refresh_entity             catalog:refresh
  ! catalog.register                   catalog:write
    platform.estimate_cost             finops:read
    platform.evaluate_policy           policy:evaluate
  ! platform.open_pull_request         delivery:propose
    platform.plan_promotion            delivery:plan
    platform.render_workload           platform:render
  ! scaffolder.execute                 scaffolder:execute
    scaffolder.get_template_parameters scaffolder:read
    scaffolder.list_templates          scaffolder:read

  withheld from this identity:
    x platform.approve_promotion       identity lacks delivery:approve
    x platform.detect_drift            identity lacks observability:read

apiVersion: backstage.io/v1alpha1
kind: AiResource
metadata:
  name: platform-agent
spec:
  type: agent
  owner: group:default/platform-team
  identity: platform-agent
  # The blast radius: twelve entries,
  # and note which one is absent.
  allowedTools:
    - catalog.query
    - catalog.get_entity
    - catalog.refresh_entity
    - catalog.register
    - scaffolder.list_templates
    - scaffolder.get_template_parameters
    - scaffolder.execute
    - platform.render_workload
    - platform.evaluate_policy
    - platform.estimate_cost
    - platform.plan_promotion
    - platform.open_pull_request
    # platform.approve_promotion is not
    # here, and never will be.

./run.sh tools platform-agent    # 12/14
./run.sh tools drift-agent       #  5/14
./run.sh tools cost-reviewer     #  3/14
./run.sh tools release-manager   # 14/14

Three lines in. A managed, encrypted, tagged database out.

A developer writes resources: {db: {type: postgres}} and stops thinking. The platform team decides, in one file, that "postgres" means a Crossplane composite with encryption, a private endpoint, sized backups and a cost-centre tag — not an unmanaged StatefulSet running a mirrored image. Swap that file and every service in the estate gets the new definition on its next render.

resources:
  db:
    type: postgres

# That is the whole thing. The only
# infrastructure a product engineer
# at Northwind ever writes.

apiVersion: platform.northwind.io/v1alpha1
kind: SQLInstance
metadata:
  annotations:
    k8s.score.dev/resource-guid: 6b7dbc6f-1b90-faba-e51d-03539eacd311
    k8s.score.dev/resource-uid: postgres.default#payments-ledger.db
    k8s.score.dev/source-workload: payments-ledger
    northwind.io/provisioner: template://northwind/postgres-crossplane
  labels:
    app.kubernetes.io/managed-by: northwind-platform
    app.kubernetes.io/name: payments-ledger
    backstage.io/component: payments-ledger
    northwind.io/cost-center: CC-4471
    northwind.io/data-classification: pci
    northwind.io/data-residency: eu
    northwind.io/environment: prod
    northwind.io/owner: payments
    app.kubernetes.io/part-of: payments-platform
  name: nw-payments-ledger-prod
  namespace: payments-prod
spec:
  crossplane:
    compositionRef:
      name: sqlinstance.postgres.northwind.io
  parameters:
    backupRetentionDays: 30
    costCenter: CC-4471
    databaseName: payments_ledger
    engine: postgres
    engineVersion: '17.4'
    instanceClass: db.t4g.medium
    masterUsername: nw_payments_ledger
    multiAz: true
    passwordSecretRef:
      key: password
      name: nw-db-6b7dbc6f-1b90-faba-e51d-03539eacd311
    publiclyAccessible: false
    region: eu-west-1
    storageEncrypted: true
    storageGb: 500

The part that actually kills platforms.

Golden paths get a service created. What kills platforms is month four, when the estate has quietly diverged from what git says is true and nobody knows which of the two is correct.

Nine pieces, every one replaceable.

Nothing is vendored or reimplemented. ./bin/setup.sh fetches official release artefacts at pinned versions. Where a project has no agentic story yet, this page says so rather than implying one.

Versions verified against upstream release pages in August 2026. Deliberately excluded, with reasons in the README: Cyclops and Kusion (both dormant for ~13 months), and Port (closed-source core, archived MCP server, AI agents behind a paid tier).

Acceptance checks

Produced by src/build_report.py, which executes and times each check against the real binaries and writes outputs/verify-report.json. The same suite runs in GitHub Actions on every push, plus weekly — because the failure mode worth catching is an upstream release changing behaviour under a demo that claims to be reproducible.

Four ways to run it. None of them need an account.

No cluster, no cloud credentials, no API key, no LLM. Python 3.10+ and about ninety seconds.

# what the container does before you get a prompt
pip install -r requirements.txt
./bin/setup.sh --all      # pinned upstream binaries
./run.sh test             # 120 unit tests
./run.sh verify           # 15/15

git clone https://github.com/adventurewave-labs/\
agentic-platform-engineering-extravaganza
cd agentic-platform-engineering-extravaganza

./run.sh setup   # pinned upstream binaries
./run.sh demo    # eight acts and a scorecard
./run.sh verify  # 15 acceptance checks

docker compose run --rm demo
docker compose run --rm verify
docker compose up site   # :8080
docker compose up mcp    # :8099

# stdio
claude mcp add northwind -- \
  python3 src/platform_mcp.py

# or streamable HTTP
./run.sh mcp 8099
claude mcp add --transport http \
  northwind http://127.0.0.1:8099/mcp

How this was built — no smoke, no mirrors.

A demo that overstates itself is worse than no demo. Here is the line, drawn honestly, and the command that checks it rather than asking you to take this table's word for it.