A precisely limited question
Can separate agents reach the same numerical value when each sees only a neighbour’s value? This is an agreement problem, and it is small enough to analyse directly. It is not yet a problem about intelligent decisions, correct beliefs, or biological adaptation. Keeping the question narrow is an advantage: the assumptions and the result can be checked. The example below is a mathematical demonstration, with no values inferred from shrimp observations.
Consensus theory provides a wider engineering framework for questions of this kind. Olfati-Saber, Fax, and Murray’s 2007 review connects interaction graphs with agreement dynamics and discusses the importance of information flow and system assumptions. The present note works through a deliberately simple two-agent case before considering what would have to change in a larger or biological system.
Two agents and one update rule
Let the values at step t be a(t) and b(t). Choose a fixed mixing weight α with 0 < α < 1. Both agents update simultaneously: a(t + 1) = (1 − α)a(t) + αb(t), and b(t + 1) = αa(t) + (1 − α)b(t). Each new value is a weighted average of the two previous values. The simultaneous schedule matters: both equations use values from the same old step.
Add the equations and the sum is unchanged. Subtract them and the difference d(t) = a(t) − b(t) obeys d(t + 1) = (1 − 2α)d(t). Because |1 − 2α| < 1, the difference tends to zero. A preserved sum plus a vanishing difference means both values approach their initial mean. This is a complete explanation of the result for this model, not a description of a hidden supervisor.
The boundary cases teach the lesson
Take the invented initial values 0 and 10. With α = 1/4, the first update gives 2.5 and 7.5; the second gives 3.75 and 6.25. The mean remains 5 while the gap halves each step. With α = 1/2, both reach 5 after a single update. With α = 1, they exchange values forever unless they were equal initially. With α = 0, neither changes. A small change in an assumption can change the qualitative conclusion.
These cases suggest useful verification checks for an implementation. A program should preserve the mean, reproduce the hand-calculated steps, and show the expected failure to converge at the two excluded boundary values for unequal starting states. A plot is helpful, but the calculation is the stronger test here. There is no need to select a visually smooth run from a collection of unexplained outputs.
More agents require more conditions
For a fixed finite undirected graph with symmetric, nonnegative weights aᵢⱼ = aⱼᵢ ≥ 0, connected through its positive-weight edges, define the combinatorial Laplacian L = D − A, where Dᵢᵢ = Σⱼ aᵢⱼ. Continuous-time averaging dx/dt = −Lx converges to the ordinary mean of the initial scalar states. This statement depends on those conditions. A disconnected graph permits separate components to retain different averages. Directed influence, delays, and changing contacts are different models requiring their own analysis; the two-agent calculation does not supply their guarantees.
Even where a process converges, the common value need not be useful. If all starting measurements are biased in the same direction, averaging does not remove the shared bias. If a network gives some agents disproportionate influence, agreement need not equal the ordinary average. The design question therefore has two parts: whether the system reaches agreement, and whether the agreed value is appropriate for the intended task.
What a biological connection would require
A biological application would need an observable quantity corresponding to the scalar state, evidence about which interactions convey relevant information, and a defensible account of the update rule. It would also need to explain whether continuous or discrete time is suitable and whether the network can be treated as fixed. Naming animals “agents” does not establish these correspondences. Neither does observing that a group acts together.
AESRD proposes to use examples like this to teach the discipline of modelling: define the system, derive a consequence, test the implementation, and state where the interpretation stops. The current note establishes only the behaviour of the equations it contains. Its potential value for shrimp research lies in the questions those equations help formulate, not in a claim that shrimp have already been shown to calculate an average.