Epilogue: what is still unresolved

An honest model includes its open questions. The ones that follow are not implementation details: they are decisions that do not yet have a satisfactory answer, and that will probably matter to anyone researching this kind of system.

The identity of each conversation. A business factory's conversations run on the same machine. Having each one prove who it is to the rest of the system is solved; having each one isolated from the others within that machine, especially those that read outside text, is an open question, and other questions depend on its answer.

Memory written by whoever reads outside text. A conversation that serves customers reads, by definition, text it does not control. If that conversation can write memory, a malicious text could try to stay in it. Treating memory as data reduces the risk, but isolating those conversations is a condition for closing it, and it depends on the previous question.

The cost and terms of artificial intelligence engines. Each business factory thinks with an engine its business chooses. How that engine is paid for across many businesses at once, and under what terms of use, is a question as much commercial as technical, and its answer changes with the provider.

The price of not asking for approval. Chapter 9 describes a bet: few questions, well asked, instead of an approval for each action. Whether that bet is right for every business, or only for some, is something that will be learned through use.

The board that wakes. A shared board serves to see who owes what, but it only moves the work if each event wakes whoever is waiting for it (chapter 19). How to build that bridge without losing signals, and without each conversation having to wait with the exact filter, remains open.

The central data store. Each role's data store, synchronized with a central copy (chapter 14), has been tested on one computer, with several conversations at once. Moving it to a central service with object storage, with each conversation's credential and synchronization from each business factory's machine, has not been tested yet.

One copy per person. The model of one copy of the business factory per person, over a single body of work per role (chapters 3 and 4), settles what each person sees. It remains to be seen how it behaves with many people on the same roles: how much each copy costs, and what happens when two copies ask the same role for conflicting things.

What is seen and what is not. There are improvements already thought through that are not built yet: letting people watch and take control of a browser while an agent is using it, letting them build their own screens by talking with their role, having the list of connected services show each one only once and under a simple name. None of them changes the model, but all of them change how much contact people have with the technical side, which is, after all, what this model seeks to reduce to a minimum.

To close, it is worth remembering that this book describes a model in motion. Some of its decisions changed several times while it was being built, and it is reasonable to expect some to change again. What matters is that each change has, as it has so far, a reason that can be written down.