Agentic property-management platform

§Context

Renewals and maintenance at Entrata are dense with SOPs, edge cases, and a decade of legacy data. Residents want quick, correct answers about their renewal offers, their rent and charges, and their maintenance requests, on whatever channel they already use. We set out to handle that with agents that could actually operate inside the domain, not just talk about it.

§What we built

• Led and built a multi-agent platform on AWS Bedrock AgentCore, orchestrated with LangGraph, coordinating specialized sub-agents behind a single resident conversation.
• Shipped a renewals sub-agent that answers questions on rent and charges, walks residents through multiple offers, helps them compare, and lets them accept the one they prefer, over chat, SMS, and voice.
• Delivered a second sub-agent with the maintenance team that generates maintenance work orders automatically and escalates to a property manager when a request needs a human.
• Connected the agents to our legacy PHP backend and database through an MCP server, so they read and write real system state instead of a stale copy.
• Wrapped the domain SOPs in guardrails and deterministic harnesses wherever correctness mattered more than flexibility.
• Ran the stack on GPT-4.5 through our LiteLLM proxy, with continuous prompt layering, prompt optimization, and MCP tool optimization.

§Observability and reporting

I owned the SLA and observability story: LangFuse and New Relic for tracing and drift detection, and an AWS Glue ETL pipeline that turns agent activity into performance reports, so clients can see how much human effort the agents are actually saving.

§Outcome

Residents get answers on their own channel and their own schedule, property managers get clean escalations and ready-made work orders instead of repetitive back-and-forth, and the whole system stays observable enough to trust in a domain where being wrong is expensive.