2.4Message Content

5 min read

By the end of this section you can answer

Out of everything an agent knows, what is worth sending?

Message content determines whether communication supplies information that changes a receiver’s action.

In this section you will

• Compare sending a raw observation
• Compare sending an intended action
• Construct a summary selected for the receiver
• Tie message content to the receiver’s uncertainty

Observation Messages

The obvious option: tell your partner something about the world it cannot see.

Agent 1 can read the order ticket. Agent 2 cannot.

“The next order needs soup.”

This is the most direct use of a channel, and it maps onto the asymmetry exactly. Agent 1 holds a fact, agent 2 needs it, so agent 1 says it.

Figure 1

$$\tone{comm}{m^\ag_t} = f\bigl(\tone{observe}{\obs{\ag}}\bigr)$$

$f$

whatever turns an observation into something sayable

$\obs{\ag}$

this agent’s own observation, the only thing it has to report

Report the world. The message is a function of what this agent saw.

Intention Messages

A different option, and often a better one: tell your partner what you are about to do.

“I am getting the tomato.”

Nothing about the kitchen is reported here. The message is about agent 1’s own policy, and its value is that agent 2 can now stop considering the tomato and do something else.

Key takeaway

Reporting observations addresses partial observability: you know something I cannot see.

Reporting intentions addresses interdependence: our actions collide, and one of us has to commit first.

They solve the two different problems from Chapters 1 and 2, which is why a good protocol often carries both.

Intentions are also the direct answer to the duplication failure from Chapter

1. “Both agents fetch the tomato” is unfixable from observations alone, since the kitchen looks identical to both of them and the correct action depends on a decision neither has made yet. One agent announcing its choice breaks the symmetry.

Why announced intentions are not cheating

An announced intention is not the same as seeing a partner’s action. The announcement is itself an action the sender chose, it can be wrong, and the sender is free to do something else afterwards.

Joint Actions and Policies showed that independent local policies cannot randomise which agent takes a role while guaranteeing exactly one does. An intention message is one way to be the varying signal that fixes it.

Receiver-Relevant Messages

Now the question that turns this from a taxonomy into a design problem.

Agent 1 could send its entire observation. Compare what that looks like with what agent 2 actually needs:

	Content
Full observation	position=(4,2), tomato=(3,1), onion=(7,5), order=SOUP, timer=31, board=empty, holding=none
What agent 2 needs	SOUP

Agent 2 controls the stove. It cannot reach the ingredient store, cannot chop, and has no use for agent 1’s position or the onion’s location. Of everything in that first row, one field changes what agent 2 will do.

So the design question is not “what do I know?” but:

What information does the receiver actually need in order to make a better decision?

Three tests, which are worth applying to any protocol:

Does the receiver already know it? Agent 2 can see the pot. Telling it the pot is hot is a wasted message.

Can the receiver act on it? Agent 2 cannot chop. Telling it the board is free changes nothing it can do.

Would it choose differently? If agent 2 heats the pot regardless of which dish is coming, then the dish identity, however privileged, is not worth sending.

Common pitfall

The failure this prevents is sending everything. It is tempting, it is easy to implement, and with a wide enough channel it even works.

But it wastes bandwidth in proportion to how little is relevant, it makes the protocol impossible to inspect, and it means the team has learned nothing about what matters. It also fails immediately once the channel narrows, which is the next section.

Knowledge check

Agent 2 controls only the stove. Which message is most likely to be worth sending to it?

Select one answer.

A"Soup next", which names the dish, which changes when and how agent 2 should heat the pot.B"The pot is at temperature", a clear factual report about the kitchen.C"The chopping board is free", which tells agent 2 about a part of the kitchen it cannot see.DThe full observation vector, so agent 2 can decide for itself what matters.

Check answer

Message Content Summary

• Three kinds of content: observations (what I see and you cannot), intentions (what I am about to do), and relevant information (the one fact that changes your decision).
• Observations address partial observability; intentions address interdependence. A good protocol often needs both.
• An announced intention is a chosen action, not a leak of a partner’s action, and it can serve as the varying signal that independent local policies cannot produce by themselves.
• The design question is what the receiver needs, not what the sender knows. Test each candidate message: does the receiver already know it, can it act on it, would it choose differently?
• Sending the whole observation is easy, wasteful, uninspectable, and the first thing to break when the channel narrows.