How seven agents run a building's energy operation

Each agent owns one area of the bill, and only that area.

The diagnostic framing splits a building's energy spend into five areas (Operations, Waste, Penalty, Silence, Opportunity) and assigns each to one agent, with the Dispatcher and the Analyst running as cross-cutting roles. Boundaries are enforced by each agent's tool surface, not by social convention. An Operator agent that notices a tariff anomaly does not file a tariff finding; it passes the observation to the Broker through the shared event log.

Each session is a deep run, not a query.

Loop tools answer interactive questions in under five seconds. Agents run for five to fifteen minutes per cron tick. In that window an agent refreshes context, forms hypotheses, runs tools to confirm or rule out, sizes findings in dollars, and persists a session record with its reasoning trace. The longer horizon is what makes the output worth a row in the findings ledger.

Cadence matches the period the work is meaningful over.

The Operator runs every fifteen minutes because waste shows up at that cadence. The Engineer runs daily because power signatures shift gradually. The Accountant runs monthly because energy P&L is a monthly artifact. The Consultant runs monthly because opportunity windows are quarterly. The Analyst runs hourly because rollups are cheap and the team produces material continuously. Cadences are fixed in the registry and are not facility-tunable.

All seven write to the same tables.

Findings go to the findings ledger. Work orders go to work_orders. Opportunities go to opportunities. A diagnostic, a verification, or a settlement reads from those tables regardless of which agent wrote the row. No agent has private storage. The downstream platform does not have to know which agent produced what; it only has to know what the row says.

Authority belongs to the action, not the agent.

Each thing an agent does carries its own authority level. Filing a finding is observe. Creating a work order is recommend. Dispatching a contractor is adjust. Writing a control_command is dispatch. Agents do not have global trust. Their specific actions do, and the action's authority is what gates the side effect.

Opt-in per facility.

Agents are off by default. A facility activates them by flipping facilities.settings.managed_agents_enabled. Once on, every agent runs to its own cadence without further configuration. Switching them off does not delete history; sessions and findings persist.

Observation flagged for another agent

An Operator that sees a demand-charge component in a waste event does not file the penalty finding. It logs an observation against the Broker's queue. The Broker's next run picks it up and produces the finding with the right area attribution.

Work-order completion

When the Engineer's contractor closes a work order, the closure event is read by the Operator's next session as a context update. The compressor that was losing capacity should now be running clean, and the Operator's first job is to confirm the signature returned.

Finding supersession

When two agents produce findings against the same root cause from different evidence, the later finding supersedes the earlier one via the supersedes_id column. Neither row is mutated. The customer sees the latest; the audit sees both.

Escalation routing

When an action exceeds the originating agent's authority ceiling, the Dispatcher picks it up and either routes to a human operator or, for time-critical demand windows, executes within the dispatcher's own bounded authority.

Confidence below the area's threshold

Each area carries a minimum confidence for a finding to land. The Engineer's signature diagnostics require eighty-five percent confidence; the Broker's tariff switches require ninety-five percent. Sessions that exhaust hypotheses without crossing the threshold close with an 'observed but not actionable' note.

Conflicting hypotheses across agents

When two agents conclude opposite things about the same load (the Operator says waste; the Engineer says the equipment is degraded), the Dispatcher pauses both findings and escalates to a human. Conflicting evidence is more valuable as a flag than as a vote.

Action above the agent's authority ceiling

An Engineer that would otherwise dispatch a contractor on a chiller pulldown event below ten percent of nominal capacity escalates if the facility's contract authority caps that action at Recommend. The work order is filed; the dispatch waits for human approval.

Customer-facing communication required

When a finding needs an explanation rather than just an action (a tariff switch that changes the demand-charge structure, a control change that customers may notice), the Analyst is invoked synchronously to draft the customer pack and the Dispatcher holds dispatch until the customer has seen it.

telemetry_coverage_low

An agent's session aborts when the area's evidence base has less than eighty percent coverage over the relevant window. The session row records the gap; the cron retries on its next tick. Coverage gaps are themselves a Finding the Operator may file separately if they persist.

context_incomplete

The Accountant cannot file an Operations residual without the period's bills. The Broker cannot recommend a switch without the current contract. Missing context closes the session cleanly and posts a request to the operator's queue.

hypothesis_space_exhausted

A session that runs every hypothesis in its area and clears none above threshold closes with an 'observed but not actionable' summary. The next session at the next cadence re-enters with the prior session's notes.

authority_exceeded_attempt

An action whose authority exceeds the agent's ceiling for that finding class is held. The finding is filed; the action is paused; the Dispatcher routes for human review. The platform refuses the elevation rather than degrading the action silently.

session_timeout

Sessions cap at fifteen minutes of wall-clock. A session that approaches the cap stops gracefully, persists what it has, and surfaces the unfinished work in its successor's context.

agents/operator

Every fifteen minutes. Walks every facility with managed_agents_enabled, kicks off Operator sessions in parallel batches of ten.

agents/engineer

Daily at 04:00 SGT. One session per facility. Signature diagnostics on every metered circuit.

agents/accountant

Hourly trigger evaluation, with the actual P&L assembly running on the first calendar day of each month per facility.

agents/broker

Monthly on the fifth of each month per facility. Reads the prior month's tariff and contract state.

agents/consultant

Monthly on the first of each month per facility. The slowest agent because opportunity sizing requires the most modelling.

agents/dispatch

Every fifteen minutes. Routes the queue. Demand advisory runs ahead of the facility's known demand window.

agents/reports

Hourly. The Analyst's rollup. Aggregates outputs from the other six into a coherent statement per facility.