One system. Three layers. Five specialist agents.
Cloud telemetry flows in, gets reasoned over by specialist agents sitting behind one shared intent catalog, and comes out the other side as either a proposal or — once policy and a human allow it — a safely executed change.
How data becomes a decision.
Three layers, one continuous loop: observe, reason, act.
Every recommendation is traceable to a source. Every high-impact action passes through the same approval path.
That's the design principle behind the whole platform — trusted by the engineers who have to answer for what it does, not just what it flags.
Four domains live, in one architecture, verified against real AWS accounts.
Not slides, not wireframes — working code, consolidated into a single monorepo behind one shared intent catalog instead of separate per-agent systems. 2,199+ tests pass with zero failures. 50 typed intents, 43 exposed as tools over one unified MCP server — not fragmented per agent.
Five specialists. One brain.
Each agent owns a domain an SME would otherwise need a dedicated hire for. All of them sit behind the same shared intent catalog, so a cost decision knows about an open vulnerability, and an incident knows what changed in the last deploy.
Cost & rightsizing
16 registered detectors · 9 running against real accounts today- Live against real AWS accounts: idle compute, GPU waste, non-prod resources left running 24/7, storage lacking a lifecycle policy, orphaned EBS/EIP, stale snapshots, idle RDS
- Real per-resource CloudWatch metrics and pricing feed the detectors — not a guessed percentage
- Every proposed fix generates a full safe-apply Terraform sequence (import existing state → plan → review) — apply is never auto-executed, a human runs it
- Cross-domain Reports (Markdown/JSON/PDF) with per-detector dollar impact, effort, and recommendation
- A built, tested Terraform drift/reconciliation pipeline and multi-cloud cost ingestion exist in the codebase and are being wired into the live product next
Health & incidents
Live health checks across EKS, EC2, RDS, ALB, and ECS- Real cluster listing/status checks against connected AWS accounts, not simulated
- Extended beyond Kubernetes: EC2 instance health, RDS availability, load-balancer target health, ECS service capacity
- Inventory-first scanning — probes what's actually in the account once, instead of sweeping every region per check
- Z-score anomaly detection, RCA, and runbook-driven remediation are built and tested, not yet wired into the live scanning path
Posture & vulnerabilities
102+ detectors · live AWS scanning- Checks posture against CIS, SOC 2, ISO 27001, NIST in real time against connected accounts
- Detects vulnerabilities and misconfigurations across network exposure, identity & access, data protection, logging, and hardening
- Attack-path reachability tracing — finds the chain from an internet-exposed resource to what it can actually reach
- Live scan parallelized: a full account scan went from over 9 hours to roughly 10 minutes
Audit & evidence
Maps real findings to SOC 2, CIS, NIST, ISO 27001, HIPAA, and PCI- Generates an audit-ready evidence pack per framework from whatever SecOps has actually found
- Shows exactly which controls are covered and which have a gap — not a checklist filled in by hand
- Shares findings directly with the SecOps agent, so remediation and evidence stay in sync
- Turns audit season into a non-event, not a scramble
Natural language over your whole estate
An early "Ask CloudSentri" assistant exists today — the full vision is ahead.Eventually, it learns your architecture and operational history, so anyone on the team can ask a plain-language question and get a grounded answer — not a dashboard to go interpret themselves.
96% of enterprises deploy agents. Only 12% can govern them.
OutSystems surveyed 1,900 IT leaders in 2026. PwC reports 78% plan to increase agent autonomy — but only 21% have governance models ready. CloudSentri's governance is the competitive moat.
Built on principles the best teams converge on.
The 2026 State of AI Agents industry report confirms: the dominant production pattern is single tool-use with human review — exactly CloudSentri's propose→policy-check→approve→generate pipeline.