It support · Production

Druva's multi-agent copilot for streamlined data protection using Amazon Bedrock

The problem

Enterprises managing complex data environments spend hours manually investigating backup failures across dozens of systems, and tracking the high volume of metrics needed to identify cyber threats leaves organizations vulnerable when any signal is missed.

Workflow diagram · grounded in source
1
User submits NL query
Trigger
The user interacts with the supervisor agent by submitting natural language queries related to data protection, backup management, and troubleshooting.
▾ source quote
“submitting natural language queries related to data protection, backup management, and troubleshooting”
2
Supervisor agent routes request
Routing
The supervisor agent analyzes the user's input and routes the request to the appropriate sub-agents based on the nature of the query.
▾ source quote
“The supervisor agent analyzes the user's input and routes the request to the appropriate sub-agents based on the nature of the query”
3
Dynamic API selection
Ai action
Semantic search retrieves the top K relevant APIs from Bedrock Knowledge Bases, then an LLM finalizes the API selection and required parameters.
▾ source quote
“we use semantic search to retrieve the top K relevant APIs. This semantic search capability enables the system to adapt to the specific context of each user request, enhancing the Copilot's accuracy, efficiency, and scalability. Once the appropriate APIs are …”
4
Data agent retrieves via GET APIs
Integration
The data agent retrieves relevant information from Druva's systems by interacting with the GET APIs.
▾ source quote
“The data agent is responsible for retrieving relevant information from Druva's systems by interacting with the GET APIs”
5
Help agent provides guidance
Ai action
The help agent draws upon an extensive knowledge base including API documentation, user manuals, and frequently asked questions to deliver context-specific assistance.
▾ source quote
“This agent draws upon an extensive knowledge base, which includes detailed API documentation, user manuals, and frequently asked questions, to deliver context-specific assistance to users”
6
Action agent executes via POST APIs
Integration
The action agent interacts with POST API endpoints to execute operations such as initiating a backup job or modifying data protection policies.
▾ source quote
“When a user needs to perform critical actions, such as initiating a backup job or modifying data protection policies, the action agent comes into play. This agent interacts with the POST API endpoints to execute the necessary operations”
7
Human-in-the-loop approval
Human review
Users can provide additional information or explicit approvals via the user feedback node before the copilot performs critical actions.
▾ source quote
“users can provide additional information or explicit approvals by using the user feedback node before the copilot performs critical actions”
8
Response delivered to user
Output
The user receives a clear and concise response along with relevant recommendations or guidance through the user interface.
▾ source quote
“The user receives a clear and concise response, along with relevant recommendations or guidance, through the user interface”
Reported outcome

The multi-agent copilot enables 90% of routine data protection tasks to be executed through natural language interactions and reduces average time-to-resolution for data security issues by up to 70%, accelerating backup troubleshooting from hours to minutes.

Reported metrics
Routine data protection tasks via natural language90%
Time-to-resolution reduction for data security issuesup to 70%
API selection accuracy (Nova Micro)81%
API selection accuracy (Nova Lite)88%
Show all 12 reported metrics
routine data protection tasks via natural language90%
time-to-resolution reduction for data security issuesup to 70%
API selection accuracy (Nova Micro)81%
API selection accuracy (Nova Lite)88%
API selection accuracy (Nova Pro)93%
API selection accuracy range (Haiku 3, Haiku 3.5, Sonnet 3.5)91% to 92%
average response time (Nova Pro)just over one second
latency (Sonnet 3.5)eight seconds
average output tokens (Sonnet 3.5)291
average output tokens (Nova Pro)86
end-to-end system evaluation score3.3 out of 5
evaluation questions tested in initial development11
Reported stack
Amazon BedrockAmazon Bedrock AgentCore RuntimeAmazon Bedrock AgentCore Gatewaylarge language models
◆ Does this fit your context?

Compare to your context

Tell us your scale, team, and constraints. We'll show what changes at your size, what fails at your scale, and whether this case is a fit, needs adaptation, or won't scale to you. Free demo, no signup.

Compare to your context →
~30 seconds · free
Source
https://aws.amazon.com/blogs/machine-learning/harnessing-the-power-of-generative-ai-druvas-multi-agent-copilot-for-streamlined-data-protection?tag=soumet-20
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The multi-agent copilot enables 90% of routine data protection tasks to be executed through natural language interactions and reduces average time-to-resolution for data security issues by up to 70%, accelerating back…

What tools did this team use?

Amazon Bedrock, Amazon Bedrock AgentCore Runtime, Amazon Bedrock AgentCore Gateway, large language models.

What results were reported?

Routine data protection tasks via natural language: 90%; Time-to-resolution reduction for data security issues: up to 70%; API selection accuracy (Nova Micro): 81%; API selection accuracy (Nova Lite): 88% (source-reported, not independently verified).

How is this it support AI workflow structured?

User submits NL query → Supervisor agent routes request → Dynamic API selection → Data agent retrieves via GET APIs → Help agent provides guidance → Action agent executes via POST APIs → Human-in-the-loop approval → Response delivered to user.

Related it support cases

WHAT TO DO WITH THIS

Now compare it to your context

This case is one data point. Whether its pattern fits you depends on your volumes, your stack, and your exception load — that comparison is the step no case study can do for you.

Browse the It support cases →