For an engineer at Mobileye, a typical Tuesday often began not with solving complex autonomous driving puzzles, but with a repetitive digital scavenger hunt. As the company's data collection pipelines scaled to ingest thousands of driving sessions daily, the internal support queue became a bottleneck. Engineers were drowning in routine status checks, spending their most productive hours performing a tedious ritual: identifying a session, cross-referencing it with visualization tools, verifying output values, and scrubbing through logs. This manual process required roughly 15 clicks across multiple backend systems for a single query. The result was a mounting backlog of tickets and a growing frustration among internal users who simply needed to know the state of their data.

The Architecture of Automated Support

To break this cycle, Mobileye implemented a sophisticated AI agent framework powered by Amazon Bedrock AgentCore. The scale of the problem was stark: an internal analysis revealed that 66% of all support tickets were routine status inquiries. The objective for the initial Proof of Concept (PoC) was aggressive, targeting a ticket classification accuracy of 95% and a response time of under two minutes. Achieving this required more than just a prompt; it required a bridge between the cloud and the company's secure on-premises infrastructure.

At the heart of the system is a hybrid cloud architecture. Because the internal ticketing system and sensitive data reside on-premises for security and compliance reasons, they cannot be directly accessed by public cloud environments. Mobileye solved this by deploying a data bridge within their secure on-premises zone, while leveraging AWS for serverless computing, security management, and system observability. This physical separation ensures that sensitive data never leaves the secure perimeter while still utilizing the high-performance compute of the cloud.

Access to the underlying intelligence is managed through a centralized internal LLM Gateway. This gateway acts as the sole entry point for Anthropic Claude models hosted on Amazon Bedrock, enforcing corporate governance, managing API quotas, and applying security policies. By routing all requests through this gateway rather than allowing direct API access, Mobileye prevents resource waste and ensures a consistent security posture across the organization.

To handle real-time data, the team utilized the Model Context Protocol (MCP). Rather than attempting to train the model on static datasets—which would be obsolete the moment a new driving session was uploaded—MCP allows the agent to communicate directly with the drive data platform APIs. When a ticket arrives, the agent uses MCP to fetch current session statuses, system logs, and diagnostic information in real-time. The workflow follows a strict four-stage pipeline: the system receives the ticket, the LLM Gateway validates the request, Amazon Bedrock AgentCore orchestrates the execution plan via MCP to gather data, and the model synthesizes this real-time evidence into a final response for the user.

From Single PoC to Enterprise Agent Platform

While the technical success of the PoC was immediate—hitting the 95% accuracy mark and slashing response times by 90%—the team encountered a new, organizational bottleneck. Most developers at Mobileye lacked the AWS credentials or infrastructure expertise required to deploy agents into production. The intelligence existed, but the delivery mechanism was gated by the Cloud Infra team, creating a new queue of requests that threatened to slow down the adoption of AI across other departments.

This tension led to a fundamental shift in strategy: the transition from a single-use tool to a company-wide self-service agent platform. The Mobileye Cloud Infra team built a managed internal service that abstracts the complexities of AWS. This platform allows developers to deploy production-grade AI agents without needing deep cloud expertise or individual AWS permissions. By platformizing the infrastructure, Mobileye shifted the power of deployment to the people who understand the business logic best.

This platform is not a lawless sandbox; it is built on strict enterprise standards for security and cost governance. It integrates CloudWatch Alarms to provide real-time observability, ensuring that operators are notified immediately if a code exception occurs. Furthermore, the platform tracks the resource consumption and cost of every individual agent, preventing budget overruns and providing transparency into the ROI of each automated workflow.

The impact of this shift was transformative. Previously, the process of moving an agent from a concept to a production environment involved weeks of approval cycles and manual configuration. With the self-service platform, that timeline has collapsed from weeks to minutes. This evolution represents a move away from AI as a specialized project and toward AI agents as a fundamental unit of product development. By solving the deployment friction, Mobileye has enabled multiple teams to independently iterate on agents, validating hypotheses in real-time without taxing the central infrastructure team.

This transition proves that the primary challenge of enterprise AI is rarely the model's intelligence, but rather the governance and delivery pipeline that allows that intelligence to reach the end user securely.