Orion AI

This three-agent pattern is suitable for repetitive operations workflows with predictable patterns, and is not suitable for highly variable scenarios requiring creative solutions or that cannot tolerate a 10-minute diagnosis delay.

Six months after Orion AI went live, 60% of incidents can be resolved without human intervention, and the system automatically handles more than 40 alerts per day.

Orion AI depends on Bedrock's availability; if Bedrock fails, the system falls back to manual diagnosis, and the team has prepared health monitoring and manual fallback runbooks for this purpose.

Cornerstone OnDemand uses a three-agent system called Orion AI for database operations, compressing average database diagnosis time from 45 minutes to 10 minutes, a reduction of 78%.

Orion AI was built by a three-person team in six months using Amazon Bedrock and the Strands Agents framework.

Orion AI splits database operations into three narrowly scoped specialized agents: a Diagnostic Agent, a Query Generation Agent, and a Remediation Agent.

The core design of Orion AI is that no agent directly calls another agent; all handoffs are handled by a state machine managed by Strands Agents, in order to avoid circular dependencies and state conflicts.

Orion AI tracks the completion status of each phase via a workflow ID, with state persisted to external storage, allowing recovery from the last successful phase if an agent fails or times out.

Orion AI's query security mechanism requires all generated SQL to pass through a parser that rejects DML/DDL statements, allows only read-only queries, and sets query timeouts and result row count limits.

Orion AI requires human approval for destructive remediation operations and generates a rollback procedure simultaneously when a remediation plan is produced.

Error cascading refers to a downstream agent treating the erroneous output of one agent as trusted input for further processing; one prevention method is to allow downstream agents to proactively refuse uncertain conclusions (e.g., triggering a human fallback when returned data is insufficient).

Orion AI is deployed on AWS infrastructure: Amazon Bedrock handles LLM inference using Claude models, Amazon RDS stores audit logs and workflow state, AWS Lambda executes the agents, and Amazon EventBridge handles alert ingestion.

In Orion AI, each agent runs in an independent Lambda function with its own IAM role, following the principle of least privilege: the Diagnostic Agent has read-only metrics access, the Query Agent has read-only database credentials, and the Remediation Agent has write permissions but only executes after approval.

Within six months, Cornerstone OnDemand reduced database diagnosis time from 45 minutes to 10 minutes, a reduction of 78%.

Cornerstone OnDemand built a multi-agent system called Orion AI on top of Amazon Bedrock and Strands Agents.