Responding to online reviews at scale is harder than it looks. Every response needs the right tone, has to follow compliance rules, should feel personal, and ideally shouldn’t take a human hours of work to produce. Multiply that by hundreds of reviews a week, and the math gets ugly fast.
We tried the obvious solution first: hand the job to a single AI agent. Write a prompt, feed it the review, and get a response. Simple. It mostly worked. But “mostly” isn’t good enough when a bad response can damage a property’s reputation or violate a policy. So we rebuilt the system using a multi-agent workflow, and the results have been significantly better.
Here’s how it works, and why we think more teams should be thinking this way.
What Are Multi-Agent Workflows?
A multi-agent workflow is exactly what it sounds like: multiple AI agents working together, each with a narrow, well-defined job. Instead of asking one agent to do everything, you break the task into stages and assign a specialized agent to each one.
Why does this work better than a single agent? A few reasons. Focused prompts produce focused outputs. Each agent can be evaluated and improved independently. And critically, separating tasks like “do the work” and “check the work” tends to surface mistakes that a single agent reviewing its own output will miss.
Common patterns include writer/reviewer setups, planner/executor pairs, and researcher/synthesizer chains. Our review response system uses the first of these.
Our Two-Agent System
Agent 1: The Writer
The writer has one job: draft a response to the review. It takes in the review content along with context about the business, then produces a reply.
But drafting a good review response isn’t just about writing something polite. The writer has to make real decisions on every pass:
- Should contact info be included, and if so, what kind?
- Are there business-specific instructions it needs to follow (certain phrases to avoid, a particular sign-off, escalation language for complaints)?
- What tone fits this specific review and this specific business?
We’ve put a lot of work into giving the writer the context it needs to make these calls well. But even a strong writer agent makes mistakes. Which brings us to the second agent.
Agent 2: The Reviewer
The reviewer agent has a completely different job: to evaluate whether the writer did its job correctly. It looks at the original review, the business context, and the draft response, and decides whether the draft is good enough to send.
This is where multi-agent setups really shine. Asking a single model to write something and then critique its own work is a losing game. The model tends to rationalize whatever it has just produced. Separating the roles forces a genuine second look, with the reviewer approaching the draft fresh and applying criteria without the bias of having written it.
The reviewer checks for compliance, tone, accuracy, completeness, and alignment with the specific instructions that apply to that business.
The Retry Loop
When the reviewer flags a problem, the draft goes back to the writer for another attempt. We retry up to 3 times.
Why 3? It’s a balance. One retry is often enough to catch simple issues. Two or three gives the system room to work through trickier cases. But beyond that, you’re usually burning compute on a task that the agents aren’t going to solve, and it’s time to bring in a human.
After 3 failed attempts, the response is escalated. A person takes a look, figures out what’s going wrong, and handles it directly.
Human-in-the-Loop by Design
We’re not trying to remove humans from the process. We’re trying to use their time better.
The reality of review response is that most replies are fairly routine. A thank-you for a five-star review, a professional acknowledgment of a minor complaint, a standard response to a common question. These are high-volume, low-complexity tasks, and they’re exactly the kind of work that eats up a human’s day without really needing their judgment.
What humans are genuinely good at is the edge cases. The review references a situation that the agents don’t have context for. The complaint that needs a nuanced apology. The situation where something about the tone just feels off. By automating the routine 90% and routing the hard 10% to people, we get faster response times, lower costs, and better quality on the cases that actually matter.
What We've Learned
A few things stood out as we built and refined this system.
First, the reviewer agent was the biggest unlock. Going from one agent to two wasn’t just incrementally better; it was a step change in output quality. The writer gets to focus on writing. The reviewer gets to focus on catching mistakes. Neither has to compromise.
Second, the retry loop matters more than we expected. A lot of the time, the writer produces something close to good on the first pass and nails it on the second. Without the retry mechanism, we’d be escalating far more cases to humans than necessary.
Third, clear escalation paths are essential. Knowing when to stop trying and hand things to a person isn’t a failure of the system; it’s a feature. Designing that handoff well has been one of the most impactful things we’ve done.
The Takeaway
Multi-agentic workflows are a practical way to handle tasks that are too complex for a single prompt but too repetitive for a human to do manually at scale.
If you’re building something similar, our advice is simple: start by breaking the task into stages, assign each stage to a specialized agent, and build in a clear escalation path for when things go wrong. You don’t need a dozen agents. Two well-designed ones, plus a human safety net, can take you a remarkably long way.
The One Platform for Multifamily Marketing & Leasing
Wayne joined Respage in 2015 as Chief Technology Officer, making an enormous impact on the organization from the start. In his time at Respage, he has single-handedly built a top-notch Technology department, including premier Development and QA teams. He also spearheaded the creation of the AI-Powered Chatbot, the most sophisticated chatbot offered in Multifamily. In response to the COVID-19 pandemic, Wayne personally conceived the idea for the Respage Resident Amenity Scheduler and brought it to market in a two-week timeframe. Through Wayne’s vision and leadership, Respage has been able to consistently bring the most innovative products to market. He is constantly at the forefront of what the Multifamily marketplace wants and has created the infrastructure to deliver quickly and efficiently.
Prior to joining Respage, Wayne was the Lead Developer and Software Architect at Global Healthcare Exchange for 14 years.
Wayne brings over 23 years of software development experience to his leadership at Respage.
