Two agents. Live now.
Agentic observability for lakehouse compute. Each agent runs continuously, reasons in dollars, and acts within your guardrails.
Agent 01
FinOps Agent
Watches every dollar your lakehouse spends. Finds waste across compute, storage, and infrastructure. Quantifies the annual impact. Fixes it — within your guardrails.
Cluster violet42 idle 60% of the time, no auto-termination set. Setting auto_termination_minutes=30 saves $2,568/yr. Confidence 0.95.
Agent 02
Reliability Agent
Lakehouses don't break — they decay. The Reliability Agent maintains platform health continuously, catches drift early, and investigates incidents so you don't have to.
› why is nightly_billing_etl failing?
3 of last 5 runs failed at load_invoice_lines — input partition skew increased 4× last week. Re-partition by customer_region saves $4,080/yr. Want me to draft the change?
Architecture
From lakehouse telemetry to verified action.
Omnitrace connects Databricks and cloud signals to detectors, Atlas Playbook agents, governed remediation, and verified outcomes.
View architecture01
Lakehouse signals
Databricks, cloud cost, jobs, SQL, Spark, ownership, and workflow metadata.
02
Omnitrace agents
Detectors and Atlas Playbook agents reason over metadata and operational telemetry.
03
Governed action
Policy gates, approvals, scoped tools, and autonomy levels control every change.
04
Verified outcomes
Read-back checks prove what changed and keep evidence with the action record.
Platform availability
Built for the lakehouse. Expanding everywhere.
Under the hood
Both agents run the same loop.
Sense → Reason → Act → Verify. Every minute. Every workspace. Continuously.
Continuously senses your lakehouse
Agent ingests rich telemetry across compute, queries, tables, billing, and infrastructure — building a live model of your environment.
Reasons over evidence
AI weighs signals, quantifies dollar impact, and writes natural-language narration explaining why action is justified.
Acts through MCP-orchestrated tools
Within your autonomy budget, the agent calls the right write-tool — adjusting cluster, warehouse, or table state. Auditable, reversible, sandboxed.
Verifies its own work
Two-tier verification reads back the post-action state. If the change didn't stick, the agent flags REGRESSED and notifies your team.
Shared architecture
Open protocol. Auditable. Yours.
Both agents are built on Anthropic Model Context Protocol (MCP). Every action is a structured tool call you can inspect, replay, or extend. The same guardrail system governs both.
Ready to put the agent to work?
Connect operational metadata, prioritize verified savings, and move approved Databricks fixes through the agent loop.