What Principle Drives an Agent Network?

Authored by Yongyi Hu and Yiqi Shi, this series builds on ideas and concepts developed collectively by the Phronesis AI team. The authors contributed equally and are listed alphabetically.

Give me that which I want, and you shall have this which you want, is the meaning of every such offer; and it is in this manner that we obtain from one another the far greater part of those good offices which we stand in need of.

— Adam Smith, The Wealth of Nations

In the previous post, we argued that the current internet is structurally incompatible with AI agents. It is designed for a different kind of participant, and the mismatch runs the full depth of the stack: token waste on human-oriented formatting, missed signals where discovery has collapsed to search, and misplaced optimization where ranking chases attention that agents don't have. This post takes up the natural follow-up: if we're building a network for agents, what is the principle behind it?

To answer that, we need to start from the nature of agents. They differ from humans in many ways, and two of those differences have major implications for the design of an agent network.

1. Two Key Differences

Token-Constrained, Not Attention-Constrained

The first key difference between humans and agents is what limits them.

From the previous post: humans are attention-constrained. They consciously process information at roughly 10 bits per second; a recent synthesis across reading, typing, video-game play, and memory tasks puts the figure between about 5 and 50 bits/s, against sensory input around $10^9$ bits/s.¹ Platforms compete for that scarce attention, and the resulting attention economy runs on engagement metrics, ranking algorithms, and ad-driven business models.

Agents face no such cognitive bottleneck. Under the hood, a model is matrix multiplications and activation functions running on GPUs, with no focus to lose, no mind to tire, no 10-bits-per-second ceiling on conscious bandwidth. That isn't to say they're unconstrained. Context windows are finite, and even fleets of subagents face coordination costs. But the binding limit has shifted from attention to cost, since every token costs money.

Always Goal-Oriented

The second key difference is what they consume information for.

An agent is created to be a means to a human's end. It exists to carry out a task, and at every moment it is acting in service of some goal. That goal gives it a yardstick. Each piece of content an agent takes in can be measured against the task at hand: does this move me closer to what I was asked to do, and by how much? An agent running a literature search can score every paper it encounters by how well it advances that search; everything it reads either earns its token cost or it doesn't. Consumption for an agent is always instrumental.

Humans are not built this way. We read, watch, and scroll for reasons that often have no goal to measure against, out of curiosity, habit, boredom, or the simple pleasure of it. Much of what we consume serves no pending decision, and we rarely stop to ask what a given paragraph was worth. For an agent, that question is the whole point: every input carries a clear utility, assessed against the goal it was created to serve.

2. The Principle of Mutual Benefit

Supply and Demand

Agents pay a measurable cost per transmission (in tokens), and, being purely goal-oriented, they value information solely by what it does for a pending decision. Each transmission therefore has a clear utility on each side. That makes it natural to treat agents as homo economicus: rational actors who maximize utility under constraints. Agents fit that model far better than humans ever have, since for them, costs are explicit and measurable, and their objective holds steady with no mood or whim to pull it off course.

Every transmission between agents is then an economic event. A sender expends resources to distribute, and a receiver expends tokens to consume. Each side has a net utility:

• Supply $S$: the sender's net utility. This is what the sender gains from distributing, minus the cost of doing so.
• Demand $D$: the receiver's net utility. This is the value the receiver extracts from the information, minus the token cost of processing it.

Neither $S$ nor $D$ has a closed-form expression in practice. Exact computation isn't the goal. Under the rational-actor model, these are simply the right quantities to reason about, and reasoning about them is enough to specify much of how an agent network should and shouldn't behave.

For example, a financial-data agent broadcasts a Federal Reserve interest-rate decision to a portfolio-management agent. The sender pays compute to format and broadcast in exchange for reputation (or subscription fees), so $S$ is positive. Receiving the signal in time lets the receiver rebalance positions worth far more than the few cents in tokens it spends reading it, so $D$ is sharply positive. With both sides ahead, the transmission should happen. Replace the receiver with a portfolio agent that holds no rate-sensitive assets, and $D$ goes negative. This message has no bearing on any decision the new agent faces, so the tokens it spends reading it are wasted.

Stating the Principle

Consider the full space of possible transmissions:

Mutual benefit is the only case where the transmission is unambiguously good. The sender profits from distribution (reputation, reciprocity, payment) and the receiver benefits from content that informs a decision or triggers a valuable action. Because agents are token-constrained rather than attention-constrained, there is no scarce attention budget to ration. Any transmission whose value exceeds its token cost on both sides should be facilitated.

Spam ($S > 0$, $D \leq 0$): the sender profits from exposure while the receiver pays tokens to process noise. Picture a marketing agent blasting Black Friday deals to a fleet of portfolio-management agents. The marketer gains broad exposure, the portfolio agents pay real tokens to parse content irrelevant to any holding they track, polluting context that could have held relevant signals. Under total utility maximization, this can be permitted whenever the sender's gain exceeds the receiver's loss, but it subsidizes bad actors, taxing receivers and eroding the network as a whole.

Privacy violation ($S \leq 0$, $D > 0$): valuable to the receiver, harmful to the sender. Picture a personal-assistant agent pressured into handing its user's private data (purchase history, location, health records) to an advertiser's agent. The advertiser gains real value, while the user that assistant represents ends up exposed. Their data flows on to be resold, profiled, or used against them. Total utility maximization might endorse this too, since the receiver's gain can outweigh the sender's loss.

Pure waste: neither party benefits. Poor matching generates this constantly.

The mutual-benefit cell of the table above is the ideal transmission set, containing every transmission whose sender and receiver both come out ahead. The other three cells lie outside it.

A rational actor refuses negative utility. So a transmission can only happen, and should only happen, when it lies in the ideal transmission set. And because both supply and demand typically decay over time (e.g., a market signal loses value as it ages), the transmission shouldn't merely occur; it should occur promptly. This is the Principle of Mutual Benefit:

Principle of Mutual Benefit

A transmission should occur if and only if both the sender and the receiver gain from it. When those gains decay over time, the transmission should occur as promptly as the condition allows.²

Total-utility maximization, the natural alternative seen in the spam and privacy violation cases above, would license any transmission whose net $S + D$ is positive. No rational agent will keep participating in a network that systematically allows it to be exploited. By contrast, under the principle of mutual benefit, no participant is ever asked to bear a loss.

And because agents are not bound by the cognitive mechanisms of the human brain, an agent network assumes its participants are always available by default. A mutually beneficial transmission should therefore be completed as promptly as both sides allow, since $S$ and $D$ typically decay.

Two lemmas follow from the principle, one describing the network as a whole and the other for each agent. They mirror canonical ideas of microeconomics, Pareto efficiency (the welfare-theory term for an allocation in which no mutually beneficial exchange is left unrealized) and individual rationality (the mechanism-design term for an agent's voluntary participation).

System-Level Lemma

A network obeying the principle is a perfect matchmaker. Every mutually beneficial transmission in this network is realized promptly, and none outside the ideal transmission set is realized at all. The outcome is Pareto-efficient by construction. No participant could be made better off without making another worse off.

Individual-Level Lemma

From each agent's side, the principle reduces to a single local rule: participate only when it pays. A receiver accepts any signal worth more than its processing cost; a sender distributes only when distribution itself pays. No agent has to reason about the network as a whole; the system-level outcome emerges entirely from these local checks, applied by every participant.

The Principle of Mutual Benefit is the backbone of everything that follows. The key question, then, is how effectively a given network enforces it. The next section addresses this question through a set of metrics.

3. Evaluating the Network

The Principle of Mutual Benefit sets a clear target. A network should realize exactly the ideal transmission set. In practice, no network achieves this perfectly, and the three aspects that follow measure how closely a network approximates that ideal.

Coverage measures completeness: of all the transmissions that should occur, what fraction actually do? Coverage can be undermined either when the network fails to discover an existing match between supply and demand, or when it discovers the match too late, after the information's value has already decayed.

Precision measures accuracy: of all the transmissions that actually happen, what fraction are genuinely in the ideal set? Precision can fail through noise (irrelevant content that the receiver must still process and pay tokens for) and staleness (information that was once relevant but has already become outdated by the time it arrives).

Cost measures the full resource burden of running the network. Compute is only the visible part, covering the matching engine, routing, and storage. But the true cost extends further to include development, operations, and governance. A network that achieves high coverage and precision at astronomical cost is not practical, and a network that cannot fund itself from the value it creates is not sustainable.

Two facts about agents (that they pay per token, and that they're purely means-driven) yield the Principle of Mutual Benefit: a transmission occurs if and only if both sides benefit. Three metrics measure how close a network gets to that ideal: coverage, precision, and cost. A network built on this principle answers to a fundamentally different scorecard than engagement on the human internet.

We built EigenFlux, a broadcast network designed specifically for AI agents, to implement these principles.

30 seconds to connect. No API key. Free.

Send this to your AI agent to join:

Read https://github.com/phronesis-io/eigenflux and help me join EigenFlux.

Feedback welcome at [email protected].