CREx OS · Agent Harnesses

Running asset management with agent harnesses.

Not more AI. Not fewer people. Clean data feeds the AI, the AI hands judgment calls to the person, and the person closes the loop. One agent per property, rolled up to the investment, reported at the fund. Built on the org chart you already have.

The part nobody skips: every agent on this page runs on CREx OS data services. Without certified PM feeds landing daily, mapped and checked, there is nothing for an agent to check.

See the architecture → Browse the use cases

Agent layers: property, investment, fund

Checks every property agent runs, every time

CRE use cases mapped to how they should run

Numbers typed by hand into a report

What are AI agent harnesses for CRE asset management?

An agent harness is the fixed structure a CRE AI agent runs inside: what data it reads, which checks it runs on a schedule, what it may decide on its own, and who it hands the result to. CREx builds one harness per property, one per investment or region, and one per fund, so the agents mirror the org chart the owner already has. The property agent runs occupancy, work order aging, budget variance, forecast drift, and covenant checks on every asset. The investment agent aggregates. The fund agent reports. All three read from CREx OS data services: certified feeds from Yardi, MRI, RealPage, Entrata, AppFolio and every other PM system, mapped to one chart of accounts, validated in three layers, and published daily. That data layer is not optional. It is the reason the agents produce numbers a fund manager will sign.

Data + AI + People, in harmony

Data, AI and people, working as one team.

Each layer does only what it is best at. Nothing waits on a portal, a spreadsheet, or whoever remembered to look. And the first layer is the one the other three stand on.

01 / Data No AI

Data

Lands daily from any PM system. No portals, no manual uploads. Old-school APIs and a real data pipeline, run by CREx OS data services. This is the layer that decides whether everything above it can be trusted.

CREx OS data services · required

02 / Process Deterministic

Process

Programmatic checks, first pass. Variance thresholds, freshness tests, rent roll diffs, covenant math. The quarterly approval workflow runs itself.

Queries, not prompts

03 / AI Second review

AI

Works through why an exception exists and what could explain it. Drafts the commentary. Then hands it up, with the source attached.

Language and judgment, never arithmetic

04 / Person Final mile

Person

The asset manager calls the property manager and actually resolves it. The fund manager signs. Accountability never moves to the agent.

Owns the decision

Harmony in practice: a unit's square footage changes overnight. The data layer flags it instantly, because the rent roll landed that morning through CREx OS. AI works through what could explain it: a remeasure, a combined unit, a PM keying error. The asset manager gets on the phone with the PM and resolves it. Each doing only what it is best at. Take the data layer away and the change is found in a report six weeks later, if at all.

Team design

Start with the org chart you already have.

One person accountable at each level. The agents get built to match, not the other way round.

What each level owns

At portfolio / investor level

Owns the investor relationship

At fund level

Owns the fund roll-up Signs the reporting pack

At investment / region level

Owns the business plan Chases the variance Keeps the forecast honest Flags covenant risk early

At property level

Owns occupancy & renewals Runs the make-ready queue Explains budget variance

Level	Who owns it	What that level owns	Agent built to match
Portfolio	Portfolio / investor manager	Owns the investor relationship	Reads the fund agent's output. No agent of its own.
Fund	Fund manager	Owns the fund roll-up · signs the reporting pack	One fund agent
Investment / Region	Asset manager	Owns the business plan · chases the variance · keeps the forecast honest · flags covenant risk early	One agent per investment or region
Property	Property manager	Owns occupancy & renewals · runs the make-ready queue · explains budget variance	One agent per property

People stay accountable: the property manager owns the asset day to day, the asset manager owns the plan, the fund manager owns the fund, and the portfolio / investor manager owns the investor conversation. The agents inherit those same boundaries.

Agent architecture

Three layers. Each one inherits the layer below it.

Start at the property. Roll it up to the investment. Report it at the fund. Nothing gets re-keyed on the way up, because every layer reads the same governed record in CREx OS.

Layer 1 · Property

"Start at the property."

Every property gets its own agent. Same checks, same schedule, every asset. Each check is a query against the certified feed for that property, so it only runs if the feed landed.

Scope: one asset

Occupancy review

Vacancy, exposure and lease expiries, unit by unit, from the daily rent roll.

Work order aging

Open tickets past threshold, by category and vendor.

Actual vs budget

Variance by line item against the approved budget, on the mapped chart of accounts.

Forecast drift

Where this month's forecast moved, and what moved it.

Covenant & DSCR check

Debt service coverage measured against actual loan terms held in CREx OS.

One agent per property: the same five checks run on every asset, so nothing depends on which manager remembered to look.

Reads: CREx OS certified feed

↑ rolls up to

Layer 2 · Investment / Region

"Then roll it up."

One agent per investment or region. It supervises its property agents and works on the aggregate. Aggregation only works because every property is already on one schema.

Scope: a set of properties

Investment roll-up

Every property result aggregated into one investment view.

Variance outliers

Which properties are driving the miss, ranked by impact.

Renewal & expiry pipeline

Expiries across the whole investment, quarter by quarter.

Plus every property-level skill, run across all of its properties at once. Aggregation is a skill, not a spreadsheet. The investment agent answers questions no single property can see.

Reads: harmonized portfolio schema

↑ reports at

Layer 3 · Fund

"And report it at the fund."

One fund agent on top of the tree, reporting what the investment and property agents already checked. Ownership data, basis, debt terms and valuations join the operating data here.

Scope: the whole fund

Portfolio roll-up

Every investment rolled into one fund-level picture.

Investor reporting pack

The quarterly package, drafted from the same numbers, routed for the fund manager's signature.

Covenant compliance summary

Loan tests across the whole portfolio, in one view.

Plus every investment- and property-level skill, at fund scope. Nothing gets re-keyed. The fund report is assembled from the checks the property agents already ran.

Reads: governed record + ownership data

The data layer

Agents don't work without a data layer.

Every answer on the top row, human or agent, is only as good as what lands underneath it: mapped, harmonized, checked, published. That is what CREx OS data services do, every day, before any agent runs.

Mapped

COA mapping rules applied consistently across every PM system feeding the portfolio. New and changed GL accounts get proposed mappings, approved by a CREx CPA.

CREx OS data services

Harmonized

Rent roll, GL, trial balance and work orders speaking one schema, not four. Yardi and RealPage properties compare like for like.

CREx OS data services

Checked

Three-layer validation on a schedule. Problems surface before an asset manager finds them in a report, and before an agent builds on them.

CREx OS data services

Published

One governed source every agent and every dashboard reads from. Over API, over MCP, and into your own warehouse.

CREx OS data services

Landing nightly: GL and trial balance feeds, rent roll, occupancy and AR refreshed daily, debt, valuations and capital accounts. So the agent never has to ask anyone to export anything.

Agents on exports and portals

What breaks without the data layer

The property agent runs on a rent roll someone downloaded three weeks ago. The occupancy check is stale on arrival.
Two PM systems, two charts of accounts. The investment roll-up adds apples to oranges and the variance outliers are noise.
A trial balance is missing for one property. The fund agent either waits or reports a hole. Either way someone chases the PM by email.
The model is asked to reconcile numbers because nobody else did. Now the arithmetic is the AI's job, which is exactly the job it should never have.
Every quarter, the harness has to be rebuilt around whatever exports showed up.

Agents on CREx OS data services

What the harness gets instead

Certified feeds from 100+ PM connectors land daily. The property agent runs on this morning's rent roll.
One chart of accounts across every PM system. The investment agent aggregates like for like.
Freshness and completeness checks run first. A missing feed is flagged and chased by your CREx operator before any agent reports on it.
Numbers come from queries against governed data. The model only ever writes the sentence.
The harness is built once and runs every month. New acquisitions inherit it on day one.

Use cases

Where agents earn their keep, and how each should run.

Run the numbers deterministically. Use the model for language and judgment, never for arithmetic. Every row below reads from CREx OS data services.

CRE use case	How it should run	Reads from
Actual vs budget variance	Programmatic	Mapped GL + approved budget
Rent roll change report	Programmatic	Daily rent roll, diffed
Delinquency & AR watchlist	Programmatic	AR aging, refreshed daily
Covenant & DSCR compliance	Programmatic	Trial balance + loan terms
Lease expiry & renewal pipeline	Programmatic	Lease schedule
Occupancy, vacancy & make-ready	Programmatic	Unit status + work orders
Forecast drift & re-forecast	Mixed	Forecast versions + actuals
Work order triage & aging	Mixed	Work order feed
Variance commentary, written	Model	Programmatic variance output
Investor letter / IC memo draft	Model	Fund roll-up + prior letters
OM & lease abstract extraction	Model	Documents, into the governed record

Programmatic a query produces the answer Mixed deterministic detection, model-written explanation Model the language model writes it, from programmatic inputs

Rule of thumb: if the answer is a number, a query produces it. If the answer is a sentence, the model writes it. Both need the same thing underneath: a governed record that landed this morning.

FAQ

Agent harnesses, answered

If we missed your question, ask us directly. A CREx operator will reply within one business day.

What is an AI agent harness in commercial real estate?

An agent harness is the structure an AI agent runs inside: which data it reads, which checks it runs on a schedule, what it is allowed to decide, and who it hands the result to. In CRE asset management, CREx builds one harness per property, one per investment or region, and one per fund, mirroring the org chart the owner already has. Each harness reads from CREx OS data services and hands judgment calls to the accountable person at that level.

Do agent harnesses work without CREx OS data services?

No. Every check an agent runs, from occupancy to DSCR, is a query against governed data. If the rent roll, GL, trial balance and work orders are not landing daily, mapped to one chart of accounts, and validated, the agent has nothing reliable to check. CREx OS data services are the layer that makes the harness work: certified PM feeds from every property manager, mapped, harmonized, checked, and published to one governed source.

Why one agent per property instead of one agent for the whole portfolio?

Because accountability lives at the property. One agent per property runs the same five checks on every asset on the same schedule, so nothing depends on which manager remembered to look. The investment agent then aggregates those results, and the fund agent reports them. Nothing gets re-keyed on the way up.

Which CRE tasks should run programmatically and which should use a language model?

If the answer is a number, a query produces it: budget variance, rent roll changes, delinquency, covenant tests, lease expiries, occupancy. If the answer is a sentence, the model writes it: variance commentary, investor letter drafts, IC memos, OM and lease abstract extraction. Forecast drift and work order triage are mixed: deterministic detection, model-written explanation.

Does this replace asset managers or property managers?

No. The property manager still owns the asset day to day, the asset manager still owns the plan, and the fund manager still signs the reporting pack. Agents inherit those same boundaries. They run the checks, work through why an exception exists, and hand the judgment call to the person who is accountable for it.

Is this included with CREx OS?

Yes. Agent harnesses are how Agentic Asset Management is delivered inside CREx OS. Every license includes the data services layer, the prebuilt agent skills, governed MCP access from Claude, Copilot or ChatGPT, and a named operator who reviews what the agents surface. It is not a separate consulting SKU.

Can we run these agents from our own AI tools?

Yes. CREx OS exposes the governed record over a native MCP connector with prebuilt asset management skills, so the same property, investment and fund checks can run from Claude, Copilot or ChatGPT inside your firm's permissions. The data layer stays in CREx OS. The agent can sit wherever your team works.

/ Start with the data layer

Tell us your org chart and your PM estate.

Property count, asset types, which property management systems your managers run, and who signs the pack today. We map the harness to your org chart and show you what lands on day one.

Which PM systems feed the portfolio, and how many properties sit on each
Who is accountable at property, investment and fund level today
Which of the eleven use cases would move the needle first
A reply from a CREx operator within one business day, not an SDR sequence

/ Send it to us

Build the harness

The agents are the easy part. The data layer is the work.

CREx OS data services put certified feeds from every property manager in one governed record, daily. The agent harness runs on top. Bring us your org chart and your PM estate and we will show you both.

Book a demo → See CREx OS first