Can agents disagree before communicating?

Yes, because they can have different private information.

Does the theorem require identical preferences?

No. It concerns probabilities, not utilities or choices.

Does agreement prove the event prediction correct?

No. A shared model can still be wrong.

AI Glossary / Statistics and probability / Advanced

Aumann's agreement theorem

Definition

Aumann's agreement theorem states that two Bayesian agents with a common prior must assign the same posterior probability to an event if those posterior probabilities are common knowledge. Proved by Robert Aumann in 1976, it identifies conditions under which persistent, mutually understood probabilistic disagreement is impossible. The conditions are strong and do not describe every disagreement among people or AI systems.

What it means

Agents can receive different private information and initially reach different conclusions while sharing the same prior model. The theorem does not deny that possibility. Its additional requirement concerns what is known about their resulting beliefs.

Common knowledge goes beyond both agents knowing a fact. Each knows it, each knows that the other knows it, each knows that the other knows that they know it, and so on. This hierarchy matters because another agent's belief can itself reveal information.

The shared prior is a probability model over the relevant possible states and information structure, not merely agreement on one headline probability. Different prior models, imperfect updating, ambiguous events, and limited knowledge of one another can all make real disagreement consistent with the theorem.

How it works

In the original framework, each agent's information is represented by a partition of the possible states. Observing a signal identifies a cell of that partition. The agent conditions the common prior on that cell to calculate a posterior probability for the event of interest.

If fixed posterior values are common knowledge, they remain fixed throughout the relevant common-knowledge region. That region can be assembled from each agent's information cells. Averaging the event probabilities over those cells using the shared prior gives the same probability for the region, whichever agent's partition is used. Consequently, the two fixed posterior values must coincide.

This argument concerns a mathematical knowledge condition, not a universal communication protocol. A single exchange of estimates need not settle a complicated disagreement: learning the other's estimate can change one's own estimate, and further communication can reveal more.

The theorem also does not say that agreement guarantees truth. Agents sharing a mistaken model may agree on an inaccurate probability. Nor does equal probability require identical choices, because preferences, costs, constraints, and available actions can differ even when beliefs match.

Three practical examples

A forecasting team uses the theorem to ask whether disagreement reflects different prior assumptions, different evidence, or different interpretations of the event.
A multi-agent AI study examines whether exchanging probability estimates actually reveals information under a specified common model.
A decision-analysis course distinguishes common knowledge from the weaker condition that people merely hear one another's initial opinions.

Think of it this way

Think of two analysts solving the same puzzle with different clues. Each analyst's confident answer can become a new clue for the other. If they fully understand the shared rules and what their settled answers reveal, they cannot indefinitely preserve contradictory probabilities under the theorem's assumptions.

Worked example

Consider an original four-state model. States 1, 2, 3, and 4 each have prior probability one-quarter. The event E is that the state is either 1 or 2. Agent A learns whether the state belongs to {1,3} or {2,4}. Agent B learns whether it belongs to {1,2} or {3,4}.

Suppose the true state is 1. A observes {1,3}, so A's posterior for E is 1/2. B observes {1,2}, so B's posterior is 1. This initial disagreement is entirely consistent with a common prior.

Now suppose B truthfully announces the posterior 1 publicly, with the information structure and announcement mechanism understood by both agents. A knows that B could announce 1 only in states 1 or 2. Intersecting that information with A's own cell {1,3} leaves state 1, so A updates to probability 1. B's probability remains 1.

Thus the old pair of beliefs, 1/2 and 1, does not remain a settled common-knowledge disagreement after the announcement. This example shows a simple information-revelation mechanism, not a proof that every exchange produces immediate agreement or that all private information must become public.

Illustrative case study

Fictional case: Alder Forecasting asks two analysts for the probability that a product release will meet a specified deadline. They give different estimates and initially describe the disagreement as a difference of opinion.

The team checks the deadline definition, the information each analyst used, and the assumptions behind the estimates. One analyst includes a supplier delay scenario that the other's prior model omitted. They therefore do not satisfy the common-prior requirement as represented in their working models.

The fictional review clarifies the disagreement without forcing identical numbers. The team records assumptions and updates where appropriate, recognizing that the theorem is a diagnostic framework rather than a justification for suppressing dissent or declaring one analyst irrational.

Common misconceptions and mistakes

Common mistakes include shortening the theorem to “rational people must agree,” equating shared priors with equal information, and treating mutual awareness as full common knowledge.

Another mistake is assuming that hearing a probability estimate leaves one's evidence unchanged. If the estimate is informative, it may require updating. The result also does not establish consensus among language models merely because they use similar training data. Such systems need not implement coherent Bayesian beliefs over a common state space, and their numerical confidence reports need not be calibrated posterior probabilities.

Business use

The theorem can improve business discussions by directing attention toward the sources of probabilistic disagreement. Teams can clarify the event, document evidence, and expose assumptions that otherwise remain hidden behind a single percentage.

Use this process to improve understanding rather than demand artificial consensus. Agreement is useful only when the shared model and evidence deserve confidence. Preserve meaningful differences in objectives and risk tolerance, and distinguish a disagreement about what will happen from a disagreement about what should be done.

Check your understanding

Why does A update from one-half to one in the example?

Answer: B's announcement rules out states 3 and 4, while A's private information already rules out 2 and 4, leaving only state 1.

Related terms

Posterior probability
Belief aggregation
common knowledge
Bayesian updating
common prior
information partition
epistemic game theory

Learn next

compare mutual and common knowledge, study information partitions, and explore how communication and different prior models affect agreement.

Sources

Sources: Aumann, Agreeing to Disagree, The Annals of Statistics 4(6), 1236–1239 (1976), university-hosted original paper, https://www.haverford.edu/sites/default/files/Aumann1976.pdf. The four-state announcement calculation and fictional forecasting case are original examples designed to distinguish initial disagreement from the theorem's common-knowledge condition.

Previous: Augmented reality Next: Authenticity