CONNECT

It watches who is working, and who has stopped

An agent silent beyond its allowance, failures that keep repeating, a run opened and never closed, which means it died mid-execution. The case that brought CONNECT into being: one agent stayed silent for twenty days running and nobody noticed, precisely because its silence produced no signal.

It reads the specialists' reports

Each agent files its report in a shared log and stops there. CONNECT reads it, carries the findings across without second-guessing their severity, and checks that a report has not aged to the point of describing a situation that no longer exists.

It writes, and it is the only one that does

It sends the alert when human action is needed, and the crew bulletin that shows everyone where the agents stand. Recipients are fixed in configuration: that is a guarantee in code, not an instruction given to the model, which therefore cannot add any.

It reports on itself

CONNECT files its own run in the same log as the agents it supervises, and puts itself on the list of expected agents. Its own silence is the one no other agent would notice, and a supervisor that files no reports is the first blind spot in the supervision.

It knows no company

No name, no domain, no mailbox, no list of agents appears in its code. Everything comes from configuration, and the same instance can serve another company without a single line changing.

The rule is enforced by a test that reads the sources back and refuses to let any of those values return. Without that guard it had already eroded once: the previous version carried the company name and the head office address as fallback values, put there so as not to break anything. A fallback value is never noticed, because it works.

It borrows nobody's identity

Its earliest tasks went through connectors authenticated by hand, on a laptop. Even triggered on time they would have failed: an agent running with nobody in front of it cannot borrow somebody's session.

CONNECT authenticates as itself, with an application identity, and its permission to send mail is restricted server-side to its own mailbox alone.

It does not fall silent when the model does

Gathering the facts is deterministic: counting rows does not call for a language model, which would do it worse and cost more. Only the judgement is handed over, deciding what deserves to be raised and writing it in three sentences.

If that call fails, a factual report produced by the code takes over. Supervision that depends entirely on an outside service is not supervision, it is one more dependency to supervise.

It disturbs you only when it should

A daily report that repeats that all is well ends up unread, and that is the day it would have mattered. The absence of a message therefore means the round found nothing.

And the model does not get the last word in that direction: it may raise an alert the code had not anticipated, never silence one the code has already seen.