6. Work as a record

A role's work is what the role has done and has in progress: the quotes it put together, the reports, the plans, the status of each open request. It is kept in the role's own place, with its complete history.

Nothing important lives only in a conversation

The idea that organizes this chapter is short: a conversation can be closed or restarted at any moment, and the work has to outlive it. Agents' conversations have a limit on what they can remember, they are compacted when they grow, and they are restarted when the version of the business factory changes. If the work lived in them, each of those events would be a loss.

That is why all of a role's conversations (the main one, the task conversations, and the routine conversations), in any person's copy, work on the same record, and each one leaves there what it produces. Each role keeps a single body of work per business, no matter which person a conversation is answering. The alternative, in which each conversation, or each copy, carries its own version of the work, was ruled out early: copies that diverge while the work is running are the source of almost every conflict.

A record with history

The record of the work is kept the way a development team keeps its code: every change is saved with who made it and when, and it is possible to see the difference from the previous version or to go back. But it is not used with the ritual of software development (branches, reviews of every change): in a business factory, each role saves directly onto the main version.

The choice has an advantage that is not obvious at first. A record with history turns trust into something verifiable: when a role says it did something, one can look at what changed, and when something goes wrong, it can be undone without depending on anyone's memory.

Few conflicts by design

When several conversations write to the same place, the risk is that they overwrite each other. Three simple rules reduce it:

  • one file per thing (a request, a record, a plan), never one large shared file;
  • fetch the latest before starting, and save as soon as the work is finished, in small changes;
  • if two changes are incompatible, nothing is overwritten: one's own version is saved apart, and whoever decides which one stays is notified.

Example. At Ópticas Horizonte, Sales' main conversation keeps each quote in its own file. A follow-up routine notes in that same file that the customer received the reminder. The salesperson, from their copy, reads the updated status without anyone telling them, and the proprietor, from hers, sees the same file.

Work, knowledge, and files

It is worth closing with the three things a business factory keeps that are often confused:

What Where Example
how the work is done knowledge, in the business definition the skill «How to put together a quote»
what was done and is in progress the role's work, with history yesterday's customer quote
the files shared with people the business factory's file archive the PDF that was sent to the customer

Starting each day from a note

Having the work live outside the conversations allows something more: starting each day with a new conversation. Before stopping, each conversation leaves a short note in its role's work: what it was doing, what the next step is, what remains open, what it expects from others and what it owes, and what it did not get to verify. The next day, only the roles that have something pending start, and they start from that note, not from yesterday's conversation.

The cycle of a conversation: from the day's note to the turn, to the commit, and back to the note.

The alternative (continuing the same conversation day after day) looks more convenient, but it accumulates context that is no longer useful and that is paid for on every turn. A well-written note carries what matters and leaves the rest behind.

Each role's data

Some roles work with data in table form (customers, transactions, readings from a system), and for that data a record of text files is not enough. One way to solve this is to give each role a data store of its own: a local database in which the role works, synchronized with a central copy that a service keeps. The role fetches the central copy, works, and uploads only what changed.

Three rules make it work with several conversations at once:

  • each table has a single owning role, the only one that writes to it; the others read it;
  • each table declares what identifies each row, so that changes can be compared row by row;
  • if another conversation changed the same row in the meantime, the change is not overwritten: it is set aside for whoever arrived second to review.

The last rule comes from a concrete test: with the opposite rule (the last one to write wins), a change was lost without anyone noticing. And the idea of keeping these tables with the same system that keeps the text files was also ruled out, because data files cannot be merged when two people change them at the same time.

To recap, the work lives in a record with history that all of the role's conversations share, and that is why no conversation is indispensable. The next chapter looks at how agents reach the business's systems.