Start here · Architecture
Architecture overview.
A 10-layer architecture: provider adapters, cloud snapshots, signal engine, reasoning, execution planning, approval/governance, audit, monitoring/drift, operational memory, ReleaseOps. Web today, desktop preview, multi-cloud expanding.
Read this if
01 · Form factor
The platform form factor
Axiom runs as two surfaces that share the same operational data:
- Web platform — available today. Browser-based operational command center, full feature set across cloud ops + ReleaseOps + topology + memory + workflows + approvals.
- Desktop application — macOS preview today. Windows Q2 2026. Linux Q3 2026. Adds local Terraform execution, OS keychain credential storage, native notifications, workstation mode (no outbound network).
- CLI binary — available now via brew/scoop/npm. Headless scans, plans, approvals, applies for CI/CD pipelines.
All three share the same operational memory and audit fabric.
02 · Layers
The 10 architectural layers
From bottom (cloud APIs) to top (release coordination):
- 01
Provider adapter layer
Read-only connectors for AWS (live), Azure (scan + topology live, reasoning Q2 2026), GCP (scan + topology live, reasoning Q3 2026). Each adapter normalizes provider-specific responses into a typed snapshot.
- 02
Cloud snapshot layer
Typed resource graph per scan — compute, storage, network, identity, observability — with provider tags and region scoping. Snapshots persist immutably and feed every downstream layer.
- 03
Signal engine
Pure functions over snapshots that produce findings: cost waste, security exposure, drift, performance, compliance gaps. Each signal is deterministic, testable, and traceable.
- 04
Reasoning layer
12-step cognitive loop (observe → interpret → reason → plan → verify → execute) that turns findings into prioritized, dependency-aware recommendations. Every step is auditable.
- 05
Execution planning
Phased Terraform / CLI generation with blast-radius classification, pre-flight snapshot capture, measured rollback RTO, and safety check pre-conditions per item.
- 06
Approval + governance
Multi-tier approval gates (low / medium / high risk). Trust Ladder model for measured autonomy escalation per action class. Cannot self-escalate.
- 07
Audit fabric
Immutable AxiomAuditEvent ledger for every action — provider mutations, approvals, rollbacks, user operations. SOC 2 / ISO 27001 control mapping built in.
- 08
Monitoring + drift
Continuous detection of out-of-band changes. Recurring workflows for cost anomaly watch, compliance sweep, post-execution verification, and rollback orchestration standby.
- 09
Operational memory
90-day persistent history of scans, recommendations, executions, approvals, outcomes. Per-service confidence calibration. Powers the agent's recalibration over time.
- 10
ReleaseOps layer
Deployment intelligence + governance above CI/CD systems (GitHub, GitLab, Azure DevOps, Jenkins, ArgoCD, ServiceNow). Shares topology, memory, reasoning, audit with cloud ops.
03 · Data flow
How data flows
A typical scan-to-execute lifecycle traces all 10 layers:
- Connector adapter (01) assumes IAM role, enumerates resources, normalizes responses
- Snapshot (02) persists the typed resource graph
- Signal engine (03) produces deterministic findings
- Reasoning (04) prioritizes findings + builds recommendations
- Execution planner (05) generates phased Terraform with rollback paths
- Approval gate (06) routes to the right approver tier
- Apply happens — audit fabric (07) records every step
- Monitoring (08) verifies the outcome and watches for drift
- Memory (09) records the outcome for future confidence calibration
- ReleaseOps (10) overlays release events onto this same operational graph
04
Multi-tenant isolation
Multi-tenant on shared infrastructure, isolated at every layer:
- Database rows are scoped to
organizationId; queries always filter through tenant scope - Operational memory is per-org; confidence in one org never influences another
- Per-connection External IDs prevent cross-tenant role assumption
- Audit log export is org-scoped at the API layer
See security model for the full isolation breakdown.
05
Deployment patterns
- SaaS multi-tenant — default; available on all tiers
- Single-tenant SaaS — Enterprise tier; dedicated database, dedicated reasoning quotas
- Customer-hosted control plane — Enterprise tier; deploys via Helm/Terraform into your own cloud; all data stays in your VPC
- Air-gapped workstation — Desktop Enterprise tier; local reasoning model + workstation mode; no outbound network whatsoever
Trust questions
What is this architecture optimized for?
Auditable autonomy — the system can act, but every step is observable, reversible, and policy-bounded.
Why 10 layers?
Each layer is independently auditable and replaceable. Signal engine can be tuned without touching reasoning. Reasoning can evolve without affecting audit.
Is the data flow safe?
Yes — read-only at provider layer; reasoning runs server-side in your tenant; execution requires approval; audit captures everything.
What does Axiom store across layers?
Snapshots, findings, plans, audit events. Never credentials, never object contents, never secret material.
What if I need air-gapped?
Desktop Enterprise tier with workstation mode + local reasoning model. No outbound network beyond direct AWS API.
How do I integrate?
REST API for read paths (operational memory, audit export, readiness scores). Webhook delivery for events. Connector framework for new providers.
Need a human?
Most flows are documented — but we'll help if anything is unclear.