Skip to content
Architecture

The three-layer brain.

Seeing is one layer. Thinking is another. Acting is a third. SentinelCloud separates them so each can be measured and replaced independently.

Perception

Multimodal ingestion. OTLP traces, JSON logs, Prometheus metrics, GitHub PR diffs, Slack threads, every stream becomes a uniform Signal.

  • · OTLP receiver
  • · Log shipper
  • · Metrics scraper
  • · GitHub webhook
  • · Slack/Teams adapter
Reasoning

A finite state machine over six agents: Analyst, Devil’s Advocate, Strategist, Tool-Call Critic, Safety, Verifier, plus a Narrator for human summaries.

  • · Analyst
  • · Devil’s Advocate
  • · Strategist
  • · Critic
  • · Safety / Compliance
  • · Verifier
  • · Narrator
Actuation

Every action goes through a tool registry with deny-by-default policy. Real cluster connectors are pluggable; demo uses faithful simulations.

  • · GitOps PR opener
  • · Cloud SDK shim
  • · Mesh weight toggler
  • · WAF rule writer
  • · Cache purger
  • · Feature-flag flipper
State machine

One run, twelve phases, every phase audited.

A run is not free-form chain-of-thought. It is a finite state machine. Every transition is a measurable, replayable event.

  1. 01INGEST

    Normalize signals; recall episodic memory.

  2. 02ANALYZE

    Analyst hypothesises root cause grounded in signals.

  3. 03DEBATE

    Devil’s Advocate disagrees by contract; groupthink penalty paid up front.

  4. 04STRATEGIZE

    Strategist proposes one action with rationale, cost delta, risk.

  5. 05CRITIC

    Tool-call critic validates the proposed action against its tool card.

  6. 06SAFETY

    Safety agent reads the constitution and lists violations or returns [].

  7. 07POLICY GATE

    Deterministic + LLM policy checks. If any clause fails, route to human review.

  8. 08VERIFY

    Verifier independently predicts the outcome; disagreement >X% blocks auto-act.

  9. 09CONFIDENCE GATE

    Fused calibrated confidence vs. risk-class threshold + blast radius.

  10. 10ACT

    Either auto-actuate or pause for human-on-the-loop with a natural-language summary.

  11. 11VERIFY OUTCOME

    Compare predicted vs. observed KPIs; adjust calibration.

  12. 12LEARN

    Episode + rejected alternatives + quality score persisted to memory.

The honest map

Twelve gaps. Twelve modules.

Each gap is a documented criticism of current AIOps work. Each module is the file that closes it.

IDGapModule
G1Single-LLM hallucinated commandsTool Selector Critic
G2Brittle root-cause analysisTopology-Aware Reasoner
G3Debate collapses to consensusAdversarial Debate
G4No blast-radius awarenessBlast Radius Calculator
G5No counterfactual reasoningCounterfactual Memory
G6Regex-only policy gatesSemantic Policy Engine
G7Non-reproducible benchmarksDeterministic Scenario Engine
G8FinOps ignores eviction riskCost-Risk Optimizer
G9No learning loopEpisodic Memory + PRM
G10Uncalibrated confidenceConfidence Calibration Gate
G11Reactive securityWAF Rule Synthesizer
G12Single-modal ingestionMultimodal Ingestor