Hi everyone! I need some advice on building an AI chatbot for real estate that helps users through the whole property buying process. The bot needs to handle three main phases:
Finding out what users want - asking about budget, preferences, and things they don’t want.
I’m testing both approaches for a property management client right now. Single agents work fine for simple stuff, but real estate has too many edge cases. Users want multiple things at once - like someone viewing a condo who needs mortgage rates AND neighborhood schools AND HOA fees all at the same time.
Multi-agent handles this better, but the handoffs get messy without proper state management. What saved my project was building a hybrid - one main conversational agent that spawns task-specific agents on demand instead of predetermined phases. User asks about financing while browsing? Finance agent spins up in parallel, shares context, then disappears when done. Way more flexible than rigid phase transitions.
I’ve been using LangGraph for about six months. Steep learning curve at first but it pays off for complex conversations. Their memory persistence between sessions is crucial for real estate since decisions take weeks.
One gotcha nobody warns about: real estate queries are super context-heavy, so design your agents to be stateless wherever possible or token costs explode fast.
Had the exact same problem building a real estate assistant last year. Started with a single agent thinking it’d be simpler, but the AI kept getting confused - it’d jump between phases randomly or try showing listings before understanding what the user actually wanted. Switched to a coordinator setup with specialized agents and it made a huge difference. The coordinator handles routing and uses a shared context object so agents only see the history they need. Way more reliable and uses fewer tokens too. Tried building custom orchestration first but ended up going with LangGraph - their state management is solid for tracking user preferences across different phases. Pro tip: tell users when you’re switching phases. People really like knowing when you’re moving from search mode to comparison mode.
I’ve been building conversational AI for three years now, and honestly? Start with that chain of command approach you mentioned. Here’s what I learned the hard way: real estate users are all over the place - one minute they’re making emotional decisions, the next they’re crunching numbers and comparing properties. A dialogue agent that gets user intent, plus separate agents for search and analysis, handles these mood swings way better than trying to do everything with one agent or some coordinator setup. The dialogue agent keeps the conversation flowing while your specialized agents stick to what they’re good at. One more thing - use explicit state machines instead of trying to guess what phase the user is in. People actually like knowing where they are in the process, and they can easily jump back when they inevitably change their mind about what they want.
Real estate chatbots are a nightmare because users change their minds constantly. I’ve watched too many projects crash when developers overthink architecture instead of fixing what actually breaks.
Token costs will destroy your budget with multi-agent setups. Every handoff means context switching, and real estate talks drag on when people compare properties.
What works: automate the entire flow upfront. Map every user journey and decision point, then build workflows that handle transitions automatically. No guessing about phases or routing logic.
I built one for a client - the game changer was treating it like workflow automation, not AI architecture. User mentions budget? Trigger property search. Brings up mortgage? Launch calculator. Asks about visits? Schedule automation fires.
AI handles conversation, workflows manage business logic and data flow between functions. Way more reliable than hoping agents cooperate.
Skip custom development hell and use a platform that handles workflow orchestration. You’ll ship faster and focus on UX instead of debugging state management.
Been down this road with a fintech app that had similar complexity. Token costs and latency killed us when users wanted quick answers - not the architecture choice.
Here’s what matters: users bounce between phases constantly. They look at houses, see prices, calculate payments, change budgets, start over. Phase-based thinking breaks down fast.
I’d use multiple specialized agents but ditch the coordinator. Let each agent decide when to hand off based on user input. Search agent hears mortgage rate questions? Passes to finance agent. Way more natural than some boss AI making decisions.
For context sharing, dump everything into simple JSON that follows the conversation. Budget, preferences, viewed properties, calculations. Each agent reads what it needs, updates what it touched. Don’t overthink it.
LangGraph vs custom - I’ve used both. LangGraph saves months of debugging state management, especially when users jump around or restart conversations. Learning curve’s worth it if you want to ship fast.
One thing nobody mentions: add a fallback agent that admits when the system’s confused and asks clarifying questions. Saved our user satisfaction when smart routing inevitably failed.
the multi-agent thing is kinda too much for a real estate bot. i’ve done similar stuff and a single agent gets the job done if you set the prompts right and manage the convo well. just keep it simple, users are pretty much the same. langgraph complicates things unless ur dealing with really complex use cases.