A multi-agent system for database operations
Cornerstone OnDemand has detailed Orion AI, a multi-agent operations system that it says cut average database diagnosis time from 45 minutes to 10. The 7 October AWS case study describes a three-person team building the platform in six months with Amazon Bedrock, Strands Agents and a collection of narrowly scoped operational tools. Rather than placing every capability in one assistant, Orion uses a central orchestrator to route work to 13 domain specialists.
Three SQL Server agents illustrate the division. One diagnoses blocking chains, waits and long-running queries; another investigates root blockers; and a third queries live instances for current high-CPU and session data. Other agents cover infrastructure monitoring, operational analytics, lifecycle management, runbook retrieval and notifications. The hub has routing and control-flow tools but no domain tools of its own, limiting the context and authority carried through each decision.
Measured changes to the support workflow
Cornerstone reports that Orion collapsed more than 10 manual database-lifecycle steps into one interaction and removed a 15-minute reporting lag between site reliability and data teams. Median redundant alerts fell by 65 per cent, leaving roughly three or four of every 10 previous alerts to reach engineers. The system can identify a likely root cause, recommend a fix and prepare a Jira ticket for the correct on-call person.
Those are vendor-supplied results, not an independent benchmark, but they are specific enough to show where the value was measured. The largest gain came from coordinating existing diagnostic work: connecting to the affected database, querying system views, correlating logs and translating evidence into a hypothesis. The AI layer did not remove operational data sources; it made their use more consistent and reduced the hand-offs between teams.
Fast routing and fresh operational truth
Orion uses keyword routing for predictable requests and semantic search when the intent is ambiguous. AWS says about 80 per cent of requests take the in-memory keyword path in under a millisecond. This is a useful reminder that not every AI decision needs another model call. A deterministic fast path lowers cost and latency while the semantic fallback handles the smaller set of queries that genuinely need interpretation.
Memory is also deliberately constrained. Same-session context is stored separately from long-term recall, while an ephemeral scratchpad carries findings between agents during one task. Questions about current system state bypass conversational memory and read live tools instead. That reduces the chance that an old diagnosis or configuration value is repeated as if it were current, a particularly important safeguard in operations where conditions change by the minute.
Controls around powerful tools
The architecture mixes shared Model Context Protocol services with direct SDK or REST integrations. Custom controls validate input, block prompt and SQL injection patterns, sanitise secrets and personal information, enforce role-based access and track cost. Potentially disruptive operations pause at a confirmation gate and default to denial if no approval arrives within five minutes. Routing rules also prevent live operational questions from being answered from memory.
Australian organisations evaluating a similar approach should focus on the boundary between explanation and authority. Domain-scoped agents can improve tool selection, but they still need least-privilege credentials, test environments and complete traces. Cornerstone’s results suggest that disciplined orchestration can produce measurable operational gains. The deeper lesson is that the gains came from explicit domains, fresh data paths and safety gates, not from giving a general chatbot unrestricted access to production.
Results should be compared with a strong baseline
The before-and-after figures are most useful when paired with operational quality measures. Faster diagnosis is valuable only if the root cause is correct, recommended actions are safe and engineers do not spend the saved time correcting weak tickets. Teams replicating Orion should record false escalations, missed incidents, tool-call errors and the percentage of recommendations accepted without material revision. Those measures reveal whether the system improves engineering work or merely shifts effort to review.
The case study also suggests that a small team can assemble meaningful automation from managed components when the operational domains are already understood. The three-person delivery claim should not be read as eliminating subject-matter expertise. The agent boundaries, tools, keyword routes and safeguards encode knowledge held by database and reliability specialists. Successful deployment depends on making that knowledge explicit and maintaining it as systems, schemas and incident patterns evolve.
Cost should be measured at the completed-incident level rather than per model request. A multi-agent investigation may make several calls, retrieve memory and run tools, but still be economical if it reliably saves skilled engineering time. Conversely, cheap calls add little value when weak routing creates review work. Operational dashboards should connect token and infrastructure spending to accepted diagnoses and resolved incidents.
Evidence from routine production use, including failures and reversals, will be more informative than a successful reference demonstration and should shape every later expansion of scope.