I keep getting a version of the same question, and I keep answering it badly. What is a task, actually? How is it different from a skill? Why does it matter which catalog I'm in? So let me explain how CRE Agents is actually built. There are four layers, and once you can see them, you'll know exactly what to ask me for. Start at the bottom, with datasets. There are 27 of them. A dataset is a pipe to something real with no opinions attached: county labor markets, migration flows, traffic counts, residential permits, retail spending, rent comps, zoning, physical risk. Two new ones went live Tuesday. The first checks any US address against eight federal incentive overlays (Opportunity Zones, LIHTC qualified census tracts and difficult development areas, NMTC low-income communities, HUBZone, IRA energy communities, the federal historic tax credit, and EB-5 targeted employment areas) and hands back every match plus a stacking score. The second opens up New York City: PLUTO building facts, energy benchmarking, ACRIS transactions, tax assessments, and DOB permits, all behind one lookup. Datasets are the part your AI cannot reach on its own, no matter how well you prompt it. Next up are agent skills, and there are 186 now. A skill is methodology. It carries no data and produces no file; it teaches the model how the work is done. My favorite example is our Chart of Accounts skill, because it is simultaneously the most boring and the most valuable. It holds the method for mapping a T12's line items into standardized categories that roll up to a clean, comparable NOI, then routes to a property-type reference for that type's actual category set and rollup order. There are thirteen of those references: multifamily, single-family rental, student, senior, affordable, manufactured housing, office, retail, industrial, hotel, self storage, RV resort, and general. That knowledge is what an analyst absorbs in their first two years and never writes down. We wrote it down thirteen times. Then tasks. There are 529, and 506 of those you can ask for by name. A task is a deliverable with a defined shape, and it's the layer you actually touch. Each one declares what it needs from you, which skills it pulls, which datasets it reaches into, and what comes back. Take lease rollover exposure: give it a rent roll and a property type, optionally market rent and your own rollover assumptions, and it pulls the rollover methodology and the leasing cost reserve method, hits the benchmarks dataset, and returns an expiration schedule by year, WALT, mark to market, and re-leasing cost, in Excel. The task is the contract. Skills and datasets are what it reaches for along the way. Last, catalogs, which is the newest layer and the one I've become most convinced by. A catalog is a set of tasks aimed at one person's job, and specifically at the one number that person is paid on. Acquisitions is built for the buy-side principal and acquisition fee revenue. Investment Sales Broker is built for commission. Asset Management is built for hold-period value. Investor Relations is built for the capital your LPs keep and the capital you raise next. The test for every task in a catalog is whether it moves that number. If it doesn't, it drops to the free baseline every seat gets, or it gets cut. That test is the whole reason a catalog beats browsing 500 tasks and hoping. Eleven catalogs exist today. Six are open for anyone to subscribe to, and five are private, built for a single firm and invisible to everyone else. We stood another one of those up this week and ported that firm's first two deliverables into it, on their templates, in their format, following their method. Their analysts don't get a generic version of their own work product. That capability is quietly becoming my favorite thing we sell. Here's the practical payoff for you. When you ask for something and what comes back is mediocre, the fix is almost always one specific layer down, and now you can tell me which. Missing or wrong numbers is a dataset problem. Right numbers with the wrong approach is a skill problem. The right approach but the wrong deliverable is a task problem. A perfectly good deliverable that just isn't what your seat does all day is a catalog problem. Send me the layer along with the complaint and I can usually turn it around in a day or two. Two other things from this week. You can now tell Vic which property types you actually work in, so you stop being offered tasks for asset classes you never touch. And there are connect paths for Perplexity, Manus, and Codex now, because several of you asked. Next up are catalogs for data centers and for credit and special situations, both in design. Spencer P.S. If that opening question was yours, thank you for asking it twice. Bad answers on calls are how these letters get written.