Signals
Databricks + cloud
1
Billing usage
System tables
Query history
Jobs
Cluster config
Cloud cost metadata
Production architecture
From lakehouse telemetry to verified action.
Omnitrace connects Databricks and cloud signals to detectors, Atlas Playbook agents, governed remediation, and verified outcomes.
Architecture diagram
From signals to verified outcomes.
Omnitrace turns lakehouse metadata into governed agent work without requiring access to customer data.
Customer data stays outside
Platform
Omnitrace control plane
2
Evidence graph
Detector engine
Policy store
Tenant boundary
Action ledger
Verification records
Agents
Atlas Playbook
3
FinOps agent
Reliability agent
Investigator workers
Remediation workers
Verifier workers
Model boundary controls
Outcomes
Governed action
4
Owner context
Approval gates
Scoped MCP tools
Safe remediation
Read-back checks
Outcome evidence
Metadata only
Policy-gated autonomy
Cloud-hosted model option
Verified outcomes
Zone 01
Databricks + cloud signals
Operational metadata and telemetry enter through scoped connections. Omnitrace does not need customer rows, files, query results, or business records.
Billing usage
System tables
Query history
Jobs and workflows
Cluster configuration
Cloud cost metadata
Zone 02
Omnitrace control, data, and agent plane
Signals are normalized into an evidence graph where detectors, policies, ownership, and action history stay connected.
Tenant boundary
Evidence graph
Detector engine
Policy store
Action ledger
Verification records
Zone 03
Atlas Playbook + worker agents
Atlas Playbook decides what should happen next, then delegates focused tasks to worker agents under policy and autonomy controls.
FinOps agent
Reliability agent
Investigator workers
Remediation workers
Verifier workers
Model boundary controls
Zone 04
Work, remediation, and verification
Findings become owned work, approved changes, and verified outcomes instead of ending as dashboard observations.
Jira-ready context
Approval gates
Scoped MCP tools
Guarded remediation
Read-back checks
Outcome evidence
Operating loop
The platform loop is intentionally simple.
The detailed implementation can include many detectors, workers, and policy gates. The product loop stays easy to explain: sense, reason, route, act, and verify.
01
Sense operational metadata
02
Detect waste and drift
03
Build evidence and owner context
04
Select policy and autonomy level
05
Apply approved action
06
Verify the final state
Security boundary
Serious automation without expanding the data boundary.
Omnitrace is designed to do operational work from metadata and telemetry. That keeps the product focused on waste, reliability, ownership, action, and verification rather than customer business data.
Metadata-only by design
The agent uses operational metadata, configuration, cost, workflow, and telemetry signals. Customer table contents and business records stay outside the product boundary.
Cloud-bound reasoning option
Enterprise deployments can use model endpoints hosted by the customer's cloud provider so prompt context remains inside the selected cloud boundary.
Governed action path
Manual, approval-required, and low-risk automatic modes let teams decide which remediations can run and which require review.
Review the architecture with us.
Review the metadata boundary, Atlas Playbook agents, policy gates, and verification loop for your Databricks environment.