이미지: AWS ML Blog
Where repetitive tickets were eating up engineers' time
Mobileye is an autonomous-driving company that has shipped more than 230 million EyeQ system-on-chips across roughly 1,200 vehicle models. At that scale, its internal data collection and processing pipeline ingests thousands of driving recording sessions every day, generating a constant stream of status inquiries.
The problem was that each inquiry required hands-on work. Engineers had to identify the session, cross-reference visualization tools, verify outputs, and check logs before drafting an answer — a process that took an average of 15 clicks. Fully 66% of all support tickets were these routine status checks. The team had tried scripts and rule-based automation, but these couldn't handle the wide variety of ways real requests were phrased.
Design validated in a proof of concept, then moved to production
Before full rollout, the team ran a proof-of-concept phase targeting 95% ticket classification accuracy and response times under two minutes. The agent uses an Anthropic Claude foundation model, accessed through Bedrock via Mobileye's internal LLM gateway, which handles governance and quota management for model access.
The critical component was the Model Context Protocol (MCP). Through an MCP server, the agent could call the drive data processing platform's APIs in real time at inference. This let it go beyond simple classification, delivering complete answers — including session completion status, error details, recommended debugging steps, and log links — within two minutes.
Production architecture: bridging on-premises and cloud
Once the PoC succeeded, the biggest remaining challenge was that Mobileye's internal ticketing system runs entirely on-premises and cannot be accessed directly from AWS. To solve this, the team designed a hybrid architecture spanning both environments.
On the on-premises side, a local orchestrator pulls new tickets and posts the agent's generated responses back into the ticketing system. On the AWS cloud side, the AgentCore runtime executes the agent serverlessly, automatically scaling with traffic spikes without server provisioning. Agent invocation happens through a single API call. AgentCore Observability traces the entire flow — from the initial request through MCP tool calls to the final response — logging session metrics, latency, token usage, and traces. AWS Secrets Manager stores, rotates, and retrieves the credentials the agent and backend systems need, so end users never handle credentials directly.
The workflow runs in four steps. ① The local orchestrator pulls a ticket and sends it to the agent on the AgentCore runtime → ② The agent queries real-time session data through the MCP server and calls Claude via the LLM gateway to generate a complete answer with links and recommendations → ③ The local orchestrator posts the response back to the ticketing system with comments and labels → ④ AgentCore Observability logs the entire process.
The results, in numbers
Since going into production, the system has recorded the following figures:
- Overall success rate: 98% (exceeding the 95% target)
- Response time: from hours down to about 1 minute (a 90% reduction)
- Share of tickets automated: 66% of all tickets
- Monthly volume handled: 100+ tickets
Shay Margalit, a principal engineer at Mobileye, said, "AgentCore has given our engineers back hundreds of hours they used to spend on routine lookups."
From a single agent to a company-wide platform
Once the results were proven, a new problem emerged: most developers across the organization lacked AWS credentials or infrastructure access, so they couldn't deploy directly to AgentCore. To address this, Mobileye's cloud infrastructure team built an internal agent deployment platform.
On this platform, development teams only need to specify their agent code and the required AgentCore features — memory, browser tools, code interpreter, observability, gateway. The cloud infrastructure team provisions the full infrastructure behind the scenes, including AWS IAM roles, Amazon S3 storage, Amazon CloudWatch monitoring, and Amazon Cognito authentication. Developers then receive a pre-configured bedrock_agentcore.yaml file and complete deployment with a single agentcore deploy command.
Limitations and what wasn't disclosed
The figures and architecture disclosed in the source are specific to Mobileye's internal environment. The source does not detail Amazon Bedrock AgentCore's cost structure, including pricing or billing methods. Likewise, the on-premises connection method was designed around Mobileye's particular environment, and the source provides no explanation of how it might generalize to other setups.
