flwr new @aashu/antibody

Antibody: an immune system for buildings

14 AI agents each hunt one kind of building failure: battery bank (PWR), switchgear (ELEC), wall wiring (WIRE), sprinkler riser (FIRE), concrete column (STR), elevator (LIFT), rooftop chiller (HVAC), pipe leak (H2O), facade (ENV), air quality (AIR), the building management system (CYBER), emergency generator (GEN), gas and CO (GAS) and exits and fire doors (EGRESS). A coordinator ranks their reports by risk x consequence x urgency, scores building health out of 100 and alerts the facility manager.

Human-supervised. Agents fix small problems themselves in bounded loops (Tier 0-1) and re-check the result; after two failed attempts, or when building health drops below 70, they escalate. Consequential actions (shutting a valve, tripping a breaker, dispatching a contractor, messaging tenants, any change to the building control system, turning on 24/7 monitoring) are held until a person approves them in the next message of the run series.

Governed. Every agent's actions are recorded with Sentience Governor: each agent declares its objective and scope, and Governor flags high-consequence actions and actions outside an agent's lane. Approved actions carry who approved them. The Governor profiles ship inside this app (agent/governance_profiles/), so hosted runs are governed too.

Sensor readings are simulated demo data (agent/building_data.py).

Flower features used

Feature	Where
AgentApp with 14 parallel specialist agents and a coordinator	agent/agent_app.py
Flower Runtime model (FLWR_RUNTIME_BASE_URL), or a local model via ANTIBODY_MODEL	MODEL
Built-in connectors web_search, web_fetch for safety standards (run config investigate = true)	investigate()
Account connectors slack, notion for tenant reports and work orders, when bound to the run	investigate()
start_automation for round-the-clock rescans when the manager asks	investigate()
Run-series state (context.state) for health trend, held approvals and follow-up memory	load_memory(), save_memory()
Structured run events (antibody.*) for the web console in ../antibody-web	main()
Flower run config ([tool.flwr.app.config]) for scan settings on SuperGrid	configure()
Local SuperLink + Ollama for a fully on-premises run	see ../antibody-web/README.md

Files

File	What it holds
agent/agent_app.py	The swarm: decide held approvals, sense and act (agents in parallel), rank, investigate, alert (coordinator)
agent/actions.py	Every agent's tools, their simulated effects and re-checks, the fixed rules (health < 70, 2 attempts, +/-15%) and approvals
agent/specialists.py	One job description per agent. Add an entry, its readings and its actions to add a building system
agent/building_data.py	The demo building and its sensor snapshot
agent/governance.py	Sentience Governor record per agent
agent/building_domain.py	Operation type and tier for every action (one row each), shared by the Governor record and the approval gate
tests/test_actions.py	Tier rules, approvals and two-message scans with a fake model: python -m unittest discover -s tests -v

Run

uv sync
uv run flwr build
uv run flwr login supergrid
uv run flwr chat

In the chat:

/load .
Check the building.

Paste JSON to change a reading and rescan, for example:

The leak got worse. {"H2O": {"night_flow_lpm_building_empty": 14}}