How Mobileye transformed support operations using Amazon Bedrock AgentCore

Mobileye's Data Collection pipeline swallows thousands of drive-recording sessions every day, and until recently, a big chunk of the human attention going into that pipeline was wasted. Sixty-six percent of support tickets were routine status checks, the kind of thing where an engineer had to click through 15 screens across multiple backend systems just to tell someone whether their data had processed. Scripts and rule-based automation had already been tried and had already failed, because real support requests don't follow neat decision trees.

So Mobileye's engineers, led by Principal Engineer Shay Margalit, built an AI agent instead, running on Anthropic's Claude models through an internal LLM Gateway. The key piece wasn't the language model itself but the Model Context Protocol, which let the agent reach into the drive-data platform's APIs in real time — pulling session status, logs, and diagnostics while it was answering a question, rather than working off stale information. That turned it from a glorified ticket-sorter into something closer to an investigator that could tell a user their session completed, point to the exact error and a fix, or walk them through a missing submission, all in under two minutes.

The proof of concept aimed for 95% classification accuracy and sub-two-minute responses. Production beat that: 98% success rate, response times down from hours to about a minute, and over 100 tickets a month now handled without a human touching them. The five deciding factors for going with Amazon Bedrock AgentCore were serverless infrastructure, faster resolution, built-in observability, framework flexibility, and — critically — the ability to bridge Mobileye's on-premises ticketing system, which AWS couldn't reach directly, with cloud-based AI processing without opening security holes.

The architecture that resulted is a bit of a two-country arrangement. A local orchestrator on-premises extracts tickets and posts finished answers back into Mobileye's internal system, while everything resembling heavy lifting — the agent runtime, observability tracing, and Secrets Manager handling credentials so nobody touches raw keys — runs in AWS. AgentCore Observability turned out to matter more than expected, letting engineers trace exactly which MCP calls the agent made and where it stumbled, which cut debugging time considerably.

What's arguably more interesting than the support-desk win is what Mobileye did next. Most of its developers don't have AWS credentials or the patience to provision cloud infrastructure themselves, so the Cloud Infra team turned AgentCore into an internal self-service platform. Teams hand over their agent code, specify which AgentCore capabilities they need, and get a pre-configured YAML file back — deployment is a single agentcore deploy command, with IAM roles, S3, CloudWatch, and Cognito auth already wired up to company standards. A one-off proof of concept for ticket triage became, in effect, an internal PaaS for agents across the whole company.