Skip to content

AEGIS · Stage 1

CrashLoopBackOff vertical slice

Walk the thinnest defensible product path: one tenant · one agent · one model · one tool · one approval · one audit execution ID. Remediation stays held — evidence before action.

Console remains read-only. Stage-1 proves the correlation fabric — not AutoRemediate.

  1. 01

    Freeze the contract

    AEGIS Stage-1 exit: end-to-end trace and evidence complete; no secret leakage; no production write from the console.

    AEGIS freeze
  2. 02

    One tenant scope

    Nordic Federated Bank · FS Core Banking · production. Scope filters bind every tool call.

    Command Centre
  3. 03

    One read-only agent

    Kubernetes Agent 01 passport: L2 investigate, write verbs stripped, step budget enforced.

    Agent passport
  4. 04

    One local model

    Model Gateway routes to Ollama llama-3.3-70b. Sovereign Mode denies external LLM APIs.

    Model Gateway
  5. 05

    One tool

    k8s-read MCP — get/list/watch only. Delete and exec stripped at the gateway.

    Tool registry
  6. 06

    Investigate CrashLoopBackOff

    payments-auth pod after deploy v4.21 — missing AUTH_JWKS_URI. Full timeline + sealed RCA.

    Incident workspace
  7. 07

    One immutable execution ID

    AI Control Tower exec-clb-01: prompt → firewall → agent → Ollama → MCP → evidence → policy → approval hold.

    Open exec-clb-01
  8. 08

    One approval gate

    apr-clb-01 queues ConfigMap patch + rollout. Deterministic remediator does not start until approved.

    Approval queue
  • Exit criteria: complete AI path in Control Tower
  • Exit criteria: sealed RCA with rejected hypotheses
  • Exit criteria: approval pending — no silent write