An immune system for buildings
Buildings get sick.
Antibody fights back.
14 AI agents watch a building's power, wiring, water, structure and air around the clock. They catch the infection early, rank what to fix first, and keep every reading inside the building.
Demo building modelled on Stanford's Hoover Tower. Readings are simulated.
- 14
- specialist agents
- 1
- coordinator
- 0
- readings leave the site
01 · Threats
Fourteen ways a building fails quietly.
Most failures give a warning weeks ahead: a battery that runs warm, water moving at 2 a.m., a crack that grows a little every month. A yearly inspection sees one moment. Antibody watches every few minutes.
- PWRBattery bankThermal runaway
- ELECSwitchgearHot breaker
- WIREWall wiringHidden arcing
- FIRESprinkler riserClosed valve
- STRConcrete columnGrowing crack
- LIFTElevatorDoor faults
- HVACRooftop chillerEfficiency loss
- H2OPipe leakNight flow
- ENVFacadeRain intrusion
- AIRAir qualityCO₂, VOCs
- CYBERBMS controllerOpen port
- GENEmergency generatorWon't start
- GASGas and COCO build-up
- EGRESSExits and fire doorsDark exit signs
02 · Hunt
Fourteen hunters. One coordinator.
Each agent is a specialist: an electrical engineer for the switchgear, a plumber for leaks, a security analyst for the control system. It reads its own sensors and files a report with a risk, the evidence and a fix.
priority = risk × consequence × urgency
The coordinator ranks every report in plain code, not by gut feel, so every decision can be audited.
- CYBERDefault password, port open to the internet0.95
- H2O6.5 L/min flowing with the building empty0.90
- PWRCells at 41 °C, exhaust fan faulting0.85
03 · Respond
Fix the small things. Ask a human for the big ones.
The coordinator cites the governing standard, checks Slack for tenant reports and Notion for open work orders, then acts by tier.
- Tier 0CleanupStale alerts, duplicate work orders. Bots only.
- Tier 1Self-correctReset a damper and re-check. Two tries, then escalate.
- Tier 2EscalateHealth below 70 or a failed fix alerts a person.
- Tier 3Human approvalShut a valve, trip a breaker, any control-system change.
6.5 L/min at 2 a.m. (normal 0.3). Ceiling sensor on F5 wet since 01:20.
Proposed: close the F6 riser valve · Call: plumber · Tier 3, waiting for approval
04 · Network immunity
One building gets sick. Every building gets immune.
Picture a whole campus on Antibody. Every building runs its own copy on its own hardware. When one beats a failure, it shares what it learned, never the data behind it.
Sensor streams, camera frames, floor plans, control-system details.
The lesson: "this battery model with this temperature curve failed next."
Roadmap: buildings as Flower SuperNodes, lessons sent with agent.grid. Illustrated here.
05 · Live
Run it for real.
This starts a real Flower run: 14 agents check the building in parallel, then the coordinator ranks what to fix first and writes the alert. Sentience Governor records every action the agents take, and the consequential ones wait for a person to authorize them. About one to two minutes. Readings are simulated; everything the agents do is real.
06 · Built on Flower
Why this needs Flower.
14 specialists and a coordinator in one Flower app.
web_search, web_fetch, Slack and Notion for evidence.
start_automation keeps the 24/7 watch running.
Health history and follow-ups live in context.state.
Runs on Flower SuperGrid, or on a local SuperLink with a local model so the data never leaves the building.
Next: agent.grid to share lessons between buildings.
For facility managers, building owners, hospitals, data centers, campuses and insurers.