Language can look like a system of labels attached to things. In practice, its meaning is inseparable from use: who is speaking, what is happening, and what everyone understands the next move to be. That distinction becomes unusually concrete when software begins to speak back.
Rules learned by participating
Wittgenstein’s later idea of language games is useful because it moves attention away from definitions and toward activity. Asking, promising, correcting, assigning, and joking are not interchangeable packets of text. Each belongs to a practice with expectations about what follows.
Teams accumulate these expectations without writing most of them down. A status update from one person may be a request for help; the same words from another may simply close a loop. The meaning lives partly in the relationship and the history around it.
Context is social
Agent systems often treat context as a larger window of retrievable text. That helps, but it is incomplete. Context also includes authority, timing, attention, and the difference between a thought and a decision.
Designing for agent participation means making those social distinctions legible enough for software to act without flattening the team into rules. The better system is not the one that records everything. It is the one that understands which differences matter.