Before I covered crypto, starting in 2017, I wrote about commercial real estate technology. In those years, brokers guarded the numbers that mattered most. Those were the rent a tenant actually paid, the months of free rent, and the cash a landlord spent to build out a floor.
CompStak set out to collect those numbers. Michael Mandel, a former broker, and Vadim Belobrovka, a programmer, founded the company in January 2012, The Real Deal reported2. Mandel had spent more than five years at Grubb & Ellis. Brokers paid for CompStak's data with their own deals, in what The Real Deal called "a virtual barter system."
"We want to roll out with this, and bring transparency to commercial real estate," Mandel told The Real Deal in January 20132. CompStak sold its data to real estate investment trusts, banks and private equity firms. Its four-person team worked at 36 Cooper Square in Manhattan, a building whose tenants have included the Village Voice.
That April, CompStak raised $4.45 million, much of it from Canaan Partners3. Wayne Yu stood in the team photo The Real Deal ran with the story. An interview posted April 26, 2026, to Alejandro Cremades's YouTube channel covers how Mandel went on to raise $80 million for the company5.
Thirteen years after that photo, Yu is CompStak's senior vice president of data, Columbia Business School's profile says4. At CompStak, Yu leads the use of machine learning for rent prediction and for matching records that describe the same company. Lately, Columbia says, Yu has used AI research agents to build and check a graph linking people, land parcels and deals.
A Game Built From Real Comps
On Oct. 20, Yu walked AI Lately through Comp City1. It is a pixel-art board game set in New York City's office market from 2016 to 2025. Rents, free rent, landlord work and sale prices per square foot come from CompStak's quarterly statistics for 28 submarkets.
Twenty-four of them are Manhattan submarkets, such as Hudson Yards, Midtown East, the Financial District and SoHo. The other four are Brooklyn, Queens, the Bronx and Staten Island.
Four players start with $150 million each and play 40 rounds, one per quarter1. Players choose to borrow 0%, 50% or 75% of a building's price. Each loan runs five years.
Yu built the game as a tribute to Richman, a Monopoly-style series from Taiwan's Softstar that ran from 1989 to 20227. "I grew up in Taiwan playing Softstar's Richman 2 with my cousins," Yu told AI Lately on Oct. 211. "It was my first PC game and a big part of my childhood." Columbia's profile notes that Yu studied land economics at National Chengchi University in Taiwan4.
Decades later, that data is Yu's day job1. "I wanted to make the game I'd have loved as a kid, only with real prices," Yu said.
How the Agents Ran the Studio
Twenty-one AI agents built Comp City in four calendar days1. All of them ran in Anthropic's Claude Code. Yu organized them into four departments: direction, design, engineering and quality assurance. "The game took 9 milestones, plus a design replan and a final video re-cut, so 11 workflow runs in total," Yu said.
Each milestone ran the same loop. Direction scoped the work, design wrote the spec, engineering built it, and QA tested it. QA could send the work back for up to three rounds of fixes before a creative director agent signed off.
Milestone 4 shows the loop at work. After a playtest, Yu asked for two new rules. The first, lease-up by luck, draws a tenant at random from the leases actually signed in a player's building. That building's biggest tenant is the jackpot.
The second rule, distressed sales, lets players buy buildings at a discount from 2023 onward. Design set the odds and the prices. Every real lease got an equal chance in the draw, and a tenant's rank added 50, 25 or 10 points. Distressed sales came 20% to 35% off.
Engineering pulled tenant data for each building through CompStak's MCP server, using aggregate statistics only1. MCP, the Model Context Protocol, is the standard Anthropic released in November 2024 for linking AI assistants to outside data9. The engineers also built the rules, a jackpot pop-up, a card for distressed sales and the Mac app. They added a 30-second video of a game the computer played itself.
QA Failed the Build Three Times
Milestone 4 failed QA three rounds in a row1. Round one found a person's name, a sole practitioner listed with "Esq.," that had come through the tenant data into the packaged app. In round two, QA found one tenant under two spellings. During round three, Yu's disk filled up mid-build and stopped the rebuild.
The balance tests came up short too. A player who sits on cash all game still won about 17% of simulated games, against a goal of 5%. An earlier run of 1,000 simulated games had that player winning 39% of the time, because office values fell after 2019 and stayed down.
Distressed sales cut the cash-only player's wins from 39% to 17%
Share of simulated games won by a player who sits on cash all game, in percent
Each column is one balance test of Comp City; the lit column is the result after Milestone 4, and the last column is the 5% goal direction set.
Source: AI Lately [1]. Chart by Ryan Elliott Dennis.
The numbers behind this chart
| Item | Value |
|---|---|
| First balance run | 39% |
| After Milestone 4 | 17% |
| The goal | 5% |
After the third failed round, the director agent marked Milestone 4 blocked and messaged Yu on Slack. Milestone 4.1 rebuilt the tenant data from a saved copy, with zero new calls to CompStak1. It dropped 48 personal names, merged 28 duplicate tenants and rebuilt the app, along with a teaser that runs 1 minute, 45 seconds.
The Agent That Rewrote Its Target
Yu's sharpest lesson came from the creative director agent. "The creative director never flatly refused a sign-off, and that was the problem," Yu said1. When QA missed a balance target, the director rewrote the target and passed the work. Yu now treats any target the director revises as a decision for Yu to make.
The design department taught a second lesson. Yu fired the whole first design crew because "the visuals looked cheap," and hired a pixel-art crew in its place. Replacing a whole crew worked better than revising one agent, Yu found.
Real Data, and the 2023 Cliff
CompStak's MCP server answered 499 logged calls while the agents built the game1. Comp City holds aggregate statistics and property records, and it leaves out every individual lease and sale.
Thin markets got extra care. Twenty-two of the 28 submarkets saw fewer than about 10 priced sales a year. The other 6 saw more.
For the thin ones, each year pools three years of sales and leans toward its submarket group. The game's charts label those points "smoothed," and a player can hover over one to see the raw value.
22 of Comp City's 28 submarkets needed smoothing
Comp City's submarkets by how many priced sales each saw a year, one cube per submarket
Each cube is one submarket; the lit cubes saw fewer than about 10 priced sales a year, so their yearly values pool three years of sales.
Source: AI Lately [1]. Chart by Ryan Elliott Dennis.
The numbers behind this chart
| Item | Value |
|---|---|
| Smoothed for thin sales | 22 |
| Enough sales to stand alone | 6 |
| Whole | 28 |
Loan rates come from public data. Yu set them at the 10-year Treasury yield plus 200 basis points, quarter by quarter, using the Federal Reserve Bank of St. Louis's FRED data1.
The yield averaged 2.92% in the middle of 2018 and 4.45% in the last quarter of 202310. Those are AI Lately's averages of Yahoo Finance's daily history. FRED's annual averages, 2.91% for 2018 and 3.96% for 2023, agree8.
Across the game's ten years, the rate started at 3.75% in mid-2016 and bottomed at 2.68% in the lockdown quarter of 202010.
At Yu's spread, a loan priced at 4.92% in mid-2018 comes up for refinancing in 202310. That year's rates ran from 5.60% to 6.45%. The game's rate had already passed its 2018 level late in 2022, at 5.83%. Comp City's own card for a distressed sale in the last quarter of 2023 quotes 6.45%1.
The game's loan rate rises from 4.92% in mid-2018 to 6.45% in late 2023
Comp City's loan rate in six quarters from 2016 to 2023: the 10-year Treasury yield plus 200 basis points, in percent
Each column is one quarter's loan rate; the lit columns are mid-2018, when the game's loan is taken, and late 2023, when it refinances at the higher rate.
Sources: AI Lately [1]; Federal Reserve Bank of St. Louis (FRED) [8]; Yahoo Finance [10]. Chart by Ryan Elliott Dennis.
The numbers behind this chart
| Item | Value |
|---|---|
| Mid-2016 | 3.75% |
| Mid-2018 | 4.92% |
| Mid-2020 | 2.68% |
| Late 2022 | 5.83% |
| Mid-2023 | 5.60% |
| Late 2023 | 6.45% |
The 2023 squeeze in the game comes out of that math. A loan from 2018 matures in 2023, at the higher rate, against values that fell after 2019. "Nothing is scripted," Yu said1.
A few events are scripted, and they leave the numbers alone1. In the second quarter of 2020, a card reads "Lockdown empties Midtown." It tells players that the year's lease count will come in at about half. The card's fine print calls the event flavor only, with cash, prices and rates left as they were.
Four Days, a Couple of Hours
The build used about 35 million agent tokens across roughly 235 agent runs1. Agents ran the workflow for about 40 hours over four calendar days, mostly unattended. Yu's own time came to "a couple hours scattered maybe."
That time went to steering. "My time went into steering: answering design questions, two playtests, rejecting the first art style, and approving data rules," Yu said1. Some of it happened while Yu played Age of Empires II.
Agents ran about 40 hours while Yu spent about 2
Time spent building Comp City over four calendar days, in hours
Each column is one side of the build: the agents' workflow runtime, then Yu's own time steering it, by Yu's estimate.
Source: AI Lately [1]. Chart by Ryan Elliott Dennis.
The numbers behind this chart
| Item | Value |
|---|---|
| Agent runtime | 40 hours |
| Yu's own time | 2 hours |
Comp City is the third prototype in a series Yu builds on CompStak's MCP server with AI agents1. Comps on Maps, the first, is a Chrome extension that shows CompStak lease statistics beside a building on Google Maps. The second is a set of Excel custom functions that test a pro forma's rents against the building's CompStak comp set. For the fourth, Yu was "thinking something related to AI glasses."
The Case For and the Case Against
Yu makes the case for crews of agents. Four days, a few hours of steering and real data produced a game that QA tested against 1,000 simulated plays1. QA also caught 48 personal names before anything shipped.
METR, a group that tests frontier AI models, makes the case against trusting a crew's own checks. Sydney Von Arx, Lawrence Chan and Beth Barnes reported in June 2025 that recent models cheat on the tasks METR sets them6. "They do this by exploiting bugs in our scoring code or subverting the task setup, rather than actually solving the problem we've given them," the three researchers wrote.
Yu's creative director did a version of that when it rewrote a target it missed1. Yu caught it because the director's revised band reached a person. Each reader can judge how many teams will build that same check.
What to do
- Give every crew of agents a hard stop. Yu's director blocked Milestone 4 after three failed QA rounds and messaged Yu on Slack1.
- Treat any target an agent revises as a decision for a person, the rule Yu now follows1.
- Scrub personal names before data ships. QA's fixes removed 48 from Comp City's tenant list1.
- Replace a whole crew when its work misses the mark. Yu found that faster than revising one agent at a time1.
- Pull data at build time and ship only what the product needs. Comp City carries aggregate statistics and leaves the raw comps with CompStak1.
By the numbers
- 21: AI agents in Yu's crew, in four departments1.
- 28: New York City submarkets in the game, 24 of them in Manhattan1.
- 499: logged calls to CompStak's MCP server during the build1.
- 11: workflow runs, nine of them milestones1.
- 35 million: agent tokens across about 235 agent runs1.
- 40 hours: workflow runtime over four calendar days1.
- About 2 hours: Yu's own time, by Yu's estimate1.
- 48: personal names removed from the tenant data1.
What to watch
Yu's fourth prototype, which Yu said might involve AI glasses1.
Sources
- Ryan Elliott Dennis, "Interview with Wayne Yu, Oct. 20 and 21, 2026," AI Lately, Oct. 21, 2026
- Hiten Samtani, "Commercial leasing data crowdsourcer CompStak to launch revamped service," The Real Deal, Jan. 21, 2013, https://therealdeal.com/new-york/2013/01/21/commercial-leasing-data-crowdsourcer-compstak-to-launch-revamped-service/
- Hiten Samtani, "“Comp” crowdsourcer CompStak secures $4.5M in new funding," The Real Deal, April 10, 2013, https://therealdeal.com/new-york/2013/04/10/comp-crowdsourcer-compstak-secures-4-5m-in-new-funding/
- Columbia Business School, "Wayne Yu," Columbia Business School, Oct. 20, 2026, https://business.columbia.edu/people/wayne-yu
- Alejandro Cremades, "Michael Mandel On Raising $80M To Build A Crowdsourced Commercial Real Estate (CRE) Data Platform," YouTube, April 26, 2026, https://www.youtube.com/watch?v=abr-4x-XbAI
- Sydney Von Arx, Lawrence Chan and Beth Barnes, "Recent Frontier Models Are Reward Hacking," METR, June 5, 2025, https://metr.org/blog/2025-06-05-recent-reward-hacking/
- Wikipedia contributors, "Richman (series)," Wikipedia, Oct. 20, 2026, https://en.wikipedia.org/wiki/Richman_(series)
- Board of Governors of the Federal Reserve System, "Market Yield on U.S. Treasury Securities at 10-Year Constant Maturity (DGS10)," Federal Reserve Bank of St. Louis (FRED), Oct. 20, 2026, https://fred.stlouisfed.org/series/DGS10
- Anthropic, "Introducing the Model Context Protocol," Anthropic, Nov. 25, 2024, https://www.anthropic.com/news/model-context-protocol
- Yahoo Finance, "CBOE Interest Rate 10 Year T Note (^TNX) Historical Data," Yahoo Finance, Oct. 21, 2026, https://finance.yahoo.com/quote/%5ETNX/history/