How We Built EIR: A Secure Healthcare Agent Fleet on Google Cloud

For the past month, our team has been building EIR: Healthcare Agent Fleet, a secure multi-agent platform for patient access, recovery, scheduling, and hospital operations. We created this article specifically for the purpose of entering the All Things Agentic Hackathon. Our goal was to move beyond a single healthcare chatbot and build a real fleet of agents that can collaborate over time, remember context, interact with hospital systems, and stop when human approval is required. The idea Hospitals already use many software systems, but workflows are still fragmented. A patient may need to schedule an appointment, recover after discharge, answer a follow-up, or reach a clinician. At the same time, hospital teams manage reviews, inventory, scheduling, and supply operations. EIR coordinates these workflows through specialized agents rather than asking one model to do everything. Our agent architecture The Patient Access Agent runs on Google Agent Runtime using Google ADK and Gemini 3.5 Flash. We integrated: - Agent Runtime - Memory Bank - Agent Registry - Agent Identity - Agent Gateway - Model Armor - Cloud Logging, Trace, and Monitoring Memory Bank lets EIR remember safe preferences between sessions. For example, a patient can say they prefer Main Clinic and afternoon appointments, then return in a new session and have EIR use that preference when searching for availability. Agent Identity and Agent Gateway give the agent its own least-privilege identity and govern outbound calls. Model Armor adds another security layer, while backend RBAC remains the final authorization boundary. Hospital workflows EIR includes separate experiences for patients, clinicians, and operations teams. Patients can: - View appointments - Search availability - Book, reschedule, or cancel visits - Talk to EIR - Follow a recovery workflow Clinicians can: - Review their schedule - Inspect patients - Handle recovery escalations - Resolve human review items Operations teams can: - Monitor the hospital - Inspect the agent fleet - Review observability - Manage supply and replenishment workflows Scheduling is backed by FHIR R4 on Google Cloud Healthcare API. Long-running workflows use Firestore, Pub/Sub, Cloud Scheduler, and Cloud Run. Recovery and human oversight Recovery Episodes can last for days or weeks. EIR schedules follow-ups, processes structured patient responses, checks adherence and risk signals, and escalates cases when a clinician should take over. We deliberately avoid autonomous diagnosis. Critical actions use deterministic policies and human approval. The same pattern also applies to hospital supply workflows. Procurement agents may draft purchase orders, but a person must approve them before execution. Additional Google AI We also integrated Vertex AI Veo to generate short recovery education videos from already-approved care instructions. The generative model does not invent diagnoses, medications, doses, or new clinical guidance. What we learned The biggest lesson was that useful agents need much more than an LLM. Persistent memory, identity, authorization, observability, asynchronous execution, and human approval all matter. Building EIR forced us to treat agents as real software actors instead of isolated chat interfaces. Try EIR Live demo: https://eir-ui-658898892127.us-central1.run.app/ Source: https://github.com/ybedoyab/eir Built by Yulian Bedoya, Juan Quintero, Marycielo Berrio, and Alejandro Feria Gonzalez for the All Things Agentic Hackathon. Top comments (0)