An Open Question: What Does the Loop Stand On?

This one has no claim. It’s a list of questions I can’t answer yet.

The essays here so far describe a loop. Observational data records what happened, and a semantic layer says what it means. An agent reads only what has been declared. It waits in the background and speaks up when something needs a person, and a small model at the gate judges whether it should.

Every one of those pieces assumes something underneath it: a place where the organization’s knowledge and its decisions are kept, in a form both people and agents can use. I’ve been writing as if that place exists. I’ve only described one corner of it, the metric definitions.

About knowledge

A metric definition is knowledge, and it fits in a file. Much of what an organization knows doesn’t look like that. There was a launch on Tuesday. That customer is a special case. We tried this in 2023 and it failed. Where does that live? Who writes it down, and who notices when it stops being true?

When two teams mean different things by the same word, and both have good reasons, is that one definition or two? Who decides?

Much organizational knowledge is prose: documents, threads, tickets. I argued that agents should read rows with provenance, not text that might be addressing them. Can prose come inside that boundary and still be useful? Or does knowledge have to become structured before an agent may rely on it, and if so, what is lost?

About goals

A control loop steers toward something. I wrote that the person sets the goal, and left it at one sentence. Where is a goal kept? Is it data, like a definition, with an owner and a history? Who is allowed to change it, and how would an agent know it had changed?

About decisions

What counts as a decision worth recording? An agent’s judgment is easy to log, because it is already a typed value. A decision made by six people in a meeting is not. If the record holds thousands of small machine decisions and none of the large human ones, does it describe the organization at all?

How do you connect a decision to what happened next? The outcome shows up in the data eventually, but so does everything else that changed that week. A join isn’t an explanation.

What does a decision need to carry so it can be examined later? What was known at the time is different from what is true now. How much of the first has to be kept?

About the whole

Is this one system or several? Does it belong in the warehouse, beside the events, or somewhere else? Is it something a data team can own, or does it reach too far into how the organization runs?

And is any of this new? Cybernetics has been asking how organizations regulate themselves since the 1950s, and Stafford Beer built a whole model of it.¹ I don’t know yet how much of that work transfers to a setting where some of the actors are software.

I don’t have a design to offer. I’ll write again when I have a better question or part of an answer.

Footnotes

1. Beer’s Viable System Model, set out in Brain of the Firm (1972) and later books. The Wikipedia entry is a reasonable place to start. ↩