Onboarding Agent: teaching agents the questions a new hire actually has

When I started my last job, the initial days were filled with waiting. The documentation was outdated, and I lacked the required staging configuration and proper GitHub permissions. The crucial information was scattered among individuals who were occupied with other tasks, such as the senior engineer who was paired with me, who spent my second day managing a DevOps incident.

This situation was not unique to my company. It is a common challenge faced by many organizations during employee onboarding. Motivated to improve this process, I dedicated a week to each new hire and meticulously documented my actions. I observed that the onboarding process consisted of two primary components: knowledge and people, which ultimately led to increased productivity and efficient navigation within the organization.

The essence of onboarding agents lies in their ability to address questions that cannot be answered through straightforward lookups. For instance, determining whether a specific new hire can complete a particular task by a certain deadline, based on their background and existing training, or identifying the most knowledgeable person to assign to a new hire when the obvious expert is occupied, requires more than just a lookup answer.

These decisions cannot be hardcoded, as the correct answer often evolves depending on the available evidence. Recognizing this, I developed Onboarding Agent for the Taskmaster track of Google's hackathon. The system utilizes Google ADK 2.8 with Gemini 3.5 Flash as its foundation. A key design decision was to employ a collaborative coordinator instead of a traditional workflow graph.

This approach allows the agent to hold four specialists, each responsible for addressing a different aspect of the onboarding process. By consulting the most suitable specialist at runtime, the agent can make more informed decisions. The system incorporates 38 read tools, generated from each system's live OpenAPI spec, and four write actions with idempotency keys.

Additionally, it includes retry and timeout policies and three ADK plugins that enforce critical rules such as budget caps, response truncation detection, and ensuring that actions are taken only after assigning a recorded decision to the subject. The entire process is triggered by a new-hire event over Pub/Sub and runs on Cloud Run, allowing for seamless scaling to zero when not in use.

Onboarding Agent goes beyond providing reports; it actively assigns tasks, schedules meetings, and posts introductions to relevant tickets. It ensures that each action is traceable back to a recorded judgment, complete with the supporting evidence. If the agent encounters a situation it cannot resolve, it communicates this limitation and routes the question to a designated individual, rather than making an uninformed guess.

Upon arrival on day one, a new hire will find that real work has already been assigned to them, and a schedule for their first meeting has been booked. This setup mirrors the ideal onboarding experience I wish I had experienced during my own first days at a new company.