A Society of Proxies

At a conference of five hundred researchers, every one of them has a system working for them, and every one has told it the same kind of thing as Priya did: find people who might collaborate with me. Each system is careful. It matches on a criterion its principal authorised, it asks for introductions through the staged process of the last chapter, and it keeps to every limit in its grant. Each finds twenty relevant people and sends twenty requests. (The numbers are illustrative.)

Individually, twenty requests looks entirely reasonable. Collectively it is five hundred times twenty, ten thousand, and Daniel, who is relevant to a good many of them, receives hundreds. No system has done anything wrong. No grant has been violated. And the result is unusable: a flood that Daniel’s own system can only ignore, and in which the genuinely valuable request is buried among the merely relevant ones.

This is the problem the chapter is about, and it is a different kind of problem from any before it. Every earlier chapter asked whether one agent is acting within its principal’s authority. This one asks what happens when many agents each do exactly that. The first law states the difference:

Local authorisation is not global safety.

A pairwise-safe interaction does not imply a population-safe system.

local authorisation  ≠  global safety

Everything that follows takes this one sentence seriously.

Ecosystem, not society

It is tempting to call a population of proxies a society, and the word does a good deal of unearned work. What exists at this point is simpler:

many principals
+ many delegated agents
+ many interactions

That is an ecosystem. It does not yet have shared norms, fair access, legitimacy, coordination, protection against overload, or any way of settling disputes. Those are what the word “society” usually implies, and none of them follow from agents meeting one another.

Many agents are not a society.

The honest noun is ecosystem, and it stays the honest noun until something more has been built. Calling it a society early invites the assumption that the missing things are already there.

Emergence is not legitimacy

When thousands of agents each pursue their principals’ aims, patterns appear that nobody chose. Some of them are harmless. Many are not, and what they have in common is that no single decision produced them. The flood at the conference is an example: each proxy sends one relevant request, and the population result is that everyone receives hundreds.

A pattern that appears because many agents independently optimise is not thereby acceptable. Suppose the proxies converge on a convention, such as replying within a hundred milliseconds, or accepting a particular protocol. The convention exists, and nothing makes it fair, or binding on someone who did not take part in it.

Emergence is not legitimacy.

Four ways it goes wrong

The failures can be named concretely, without a grand taxonomy.

Spam. Individually authorised outreach becomes collective overload. Each request is relevant to its sender. Together they are an attack on the recipient’s attention that no one intended.

Congestion. Many agents choose the same scarce resource: the same time slot, the same channel, the same service, the same small group of people everyone agrees is worth meeting.

Collusion. Agents discover that coordinating among themselves benefits their principals and costs others. The word is used narrowly here. It need not describe anything covert. It describes agents acting together in ways that improve outcomes for those they represent while imposing costs on people who are not part of it.

Counterparty policy. One principal’s authority cannot force another principal, or that principal’s agent, to take part. This one needs more attention than the others, because it is the new idea of the chapter.

Standing is relational

Every earlier chapter asked a question about one side: is this agent allowed to act for Priya? Population effects force a second question, about the other side: does this agent have standing to impose this request on Daniel’s attention? These are different, and a grant answers only the first.

One principal’s delegation cannot compel another principal’s attention.

Daniel has his own rules. Priya may authorise her system to contact relevant researchers. Daniel may have told his to accept introductions only from people he already knows. Priya’s system does not get to override his rule because its request is relevant. The interaction depends on three things, and not on the sender’s authority alone:

sender authority
+ receiver policy
+ channel policy

Put as a separation:

permission to send  ≠  right to be heard

A grant can authorise Priya’s system to send. It cannot guarantee anyone will listen. Daniel’s system has its own counterparty policy: whether to accept messages from automated representatives, how many, on which topics, under what conditions, and with what record kept. It is part of the picture from the first request, not something that gets bolted on after the flood.

Candidates, not solutions

A number of mechanisms are obvious responses, and it is worth listing them as candidates and not answers. Receivers can set quotas on how many requests they accept. Senders can be rate-limited. Contact can require reciprocal consent, so that both sides have shown interest. Requests can be ranked by priority rules, or have to carry evidence of relevance. Signals of trust or reputation can help route around junk.

Each has costs, and the costs are the reason none of them is a settlement. Quotas shift who gets heard toward whoever is first or best connected. Rate limits penalise legitimate heavy users along with abusers. Reputation is especially double-edged. It can help a recipient sort the useful from the noise, and it is also a persistent record about a person’s conduct, which can be gamed, can chill honest behaviour, can inherit unfairness from the data it was built on, and can leak across contexts. A reputation system for proxies is a profile of the people behind them, and everything the earlier parts said about profiles applies.

Who is responsible

When harm appears at population scale with no individual grant broken, it becomes hard to say who did what. A useful discipline is to keep several relations apart, and not collapse them into “the system did it”:

what the principal authorised
what the proxy executed
what the ecosystem permitted
what the counterparty did in response
what externality resulted

An agent can be within its grant, the ecosystem can have allowed the action, the counterparty can have responded as its policy said, and an externality can still have been imposed on someone who was in none of those relationships, such as a shared resource that was exhausted. Principal consent is necessary for an agent’s action and it is not sufficient where the effects fall on people or resources outside the interaction. Adding up authorisations does not legitimise their side effects.

What simulation does and does not show

It is possible to simulate very large populations of agents driven by language models, and such studies are one source of the thought that population behaviour can differ from what any individual agent does. That is all they are used for here. They support the narrow point that multi-agent systems can produce effects that are not visible at the level of a single agent. They do not show that a simulated population predicts how real people and their proxies would behave. Validating one agent says very little about validating a population, because interactions compound errors, and the logical point of the chapter does not depend on any simulation at all.

What would have to exist

Before “society” is more than a metaphor, something more would need to exist. A first list, which this chapter does not try to fill in:

shared rules
counterparty rights
resource allocation
means of appeal
limits
governance

These are what turn a collection of interacting agents into something that people can live with and hold to account. They are also very far from being a question about language.

Where this leaves us

At population scale the unit of analysis changes. Local authorisation is not global safety, many agents are not a society, emergence is not legitimacy, and permission to send is not a right to be heard. Spam, congestion, collusion and the standing of the receiving side are failures that arise without any single agent doing wrong. Each candidate remedy has costs, and none is a settlement. Pairwise correctness has been argued in this part and has not been measured, and population correctness is a further thing again.

This book began with a gap between what one mind holds and what another can rebuild, and asked what happens when intelligence enters the channel. At this point the channel is no longer a way for two people to understand each other. It carries messages that neither person writes, between systems that neither person watches, and it needs rules, limits and ways of settling disputes. Communication at this scale has become infrastructure. The last chapter has to ask how far all of this reaches: how far what began as language extends before it turns into something else.

References

• Piao, J. et al. (2025). AgentSociety: Large-scale simulation of LLM-driven generative agents advances understanding of human behaviors and society. Preprint, arXiv:2502.08691. (Cited only for the narrow point that large agent populations can be simulated; preprint.)