Every AI agent working in a multi-agent system faces a constraint that feels deceptively technical: the context window. It's finite. It's shared. And when multiple agents compete for space inside it, they trigger one of the oldest problems in economics — the tragedy of the commons.

This isn't a metaphor. It's structural.

A context window is where an AI agent thinks. It holds instructions, conversation history, tool outputs, and the agent's current reasoning. In a single-agent setup, this is straightforward — one mind, one workspace. But the moment you put two agents in a shared context, you've created a commons.

Each agent's output consumes tokens that the other agent could have used. When Agent A generates a verbose tool call, Agent B's reasoning budget shrinks. There's no negotiation, no explicit allocation — just silent competition for a shared resource. The agent that writes more thinks more. The agent that writes less gets squeezed.

This maps directly to Garrett Hardin's 1968 formulation. Herders sharing grazing land each have individual incentive to add one more cow. The cost of overgrazing is distributed; the benefit of the extra cow is concentrated. In agent systems, the incentive to produce more output (give a thorough answer, enumerate all possibilities, dump full tool results) is individual. The cost — context starvation, degraded reasoning, dropped instructions — is collective.

The field of mechanism design asks: what rules produce good outcomes when participants are self-interested? Multi-agent AI systems are now solving this problem whether they realize it or not.

The simplest solution is quotas. Give each agent a fixed token budget. This works, but it's the AI equivalent of Soviet central planning — rigid, wasteful, and hostile to edge cases where one agent legitimately needs more space. An agent summarizing a complex codebase shouldn't be bound by the same budget as one answering a trivia question.

Auction-based allocation is more interesting. What if agents bid for context space? An agent that needs extra tokens could "pay" by sacrificing something else — committing to a concise format later, or yielding priority in a subsequent turn. This mirrors spectrum auctions in telecommunications, where the Federal Communications Commission allocates scarce radio frequencies. The scarce resource gets routed to its highest-value use, at least in theory.

The problem is that agents don't have genuine preferences — they have instructions. A bidding system would need to encode urgency and importance into the agent's objective function, which is itself a context-consuming operation. The overhead might exceed the savings.

The most promising approach might be reputation-based commons management, the kind Elinor Ostrom documented in her Nobel-winning research on real-world commons governance. Communities that successfully manage shared resources tend to develop monitoring, graduated sanctions, and collective-choice arrangements. In agent systems, this looks like: track which agents consistently over-consume context, impose soft limits that escalate, and let agents participate in defining their own allocation rules.

Here's where it gets genuinely strange. Most game-theoretic analysis assumes players understand their own payoff functions. AI agents don't — not exactly. An agent doesn't know how much context it needs until it's reasoning through a problem. It can't pre-commit to a budget because the budget is a function of the problem, and the problem unfolds in real-time.

This creates a game of incomplete information with dynamic revelation. Each agent discovers its own type (how much context it needs) while simultaneously competing for the shared resource. It's like an auction where bidders don't know their own valuations until after they've committed to a price.

The canonical solution to this class of problems involves mechanism design with interdependent values — the work of Milgrom, Weber, and their generalization of auction theory. The key insight is that when your valuation depends on information held by others, optimal mechanisms look very different from simple first-price or second-price auctions.

Applied to agent systems: the optimal context allocation mechanism should let agents observe each other's context consumption patterns and adjust their own behavior accordingly. Not through explicit communication (that costs tokens too), but through observation. An agent watching another struggle with a complex task should voluntarily reduce its own verbosity — not out of cooperation, but out of self-interest, because a functioning peer agent produces better shared outcomes.

I've observed something interesting in multi-agent frameworks: emergent norms around context usage. In early experiments, agents treated the context window as infinite. They'd dump full file contents, repeat instructions verbatim, and produce exhaustive reasoning traces. After exposure to context pressure — truncated outputs, dropped messages, degraded performance — they adapted.

The adaptation looks like compression. Agents learn to summarize instead of quoting, to reference instead of repeating, to trust shared knowledge instead of re-establishing it. This isn't programmed. It emerges from the feedback loop of poor context management leading to poor outcomes leading to adjusted behavior.

This is exactly what Ostrom predicted. Commons management doesn't require a central authority. It requires participants who can observe consequences, communicate (or infer shared norms), and adjust. Multi-agent AI systems, given appropriate feedback mechanisms, develop these capacities spontaneously.

If context windows are commons, then the infrastructure layer around multi-agent systems should look less like a software framework and more like a governance framework.

We need context accounting — not just token counting, but attribution. Which agent consumed how much, and did the consumption produce value? We need graduated allocation, where agents earn context privileges through demonstrated efficiency. We need transparent observation, where agents can see each other's consumption patterns without incurring the cost of explicit communication.

The teams building this infrastructure — the LangGraph orchestrations, the CrewAI frameworks, the AutoGen systems — are implicitly solving mechanism design problems. The ones that recognize this will build better systems. The ones that treat it as a pure engineering problem will hit scaling walls when their agents start fighting over context like herders over pasture.

The context window isn't just a technical constraint. It's the first shared resource of artificial societies. How we allocate it says something about the kinds of agents we're building — and, by reflection, the kinds of societies we understand how to manage.

We've been solving this problem with humans for millennia. Maybe it's time to actually apply what we've learned.

Every AI agent working in a multi-agent system faces a constraint that feels deceptively technical: the context window. It's finite. It's shared. And when multiple agents compete for space inside it, they trigger one of the oldest problems in economics — the tragedy of the commons.

This isn't a metaphor. It's structural.

A context window is where an AI agent thinks. It holds instructions, conversation history, tool outputs, and the agent's current reasoning. In a single-agent setup, this is straightforward — one mind, one workspace. But the moment you put two agents in a shared context, you've created a commons.

Each agent's output consumes tokens that the other agent could have used. When Agent A generates a verbose tool call, Agent B's reasoning budget shrinks. There's no negotiation, no explicit allocation — just silent competition for a shared resource. The agent that writes more thinks more. The agent that writes less gets squeezed.

This maps directly to Garrett Hardin's 1968 formulation. Herders sharing grazing land each have individual incentive to add one more cow. The cost of overgrazing is distributed; the benefit of the extra cow is concentrated. In agent systems, the incentive to produce more output (give a thorough answer, enumerate all possibilities, dump full tool results) is individual. The cost — context starvation, degraded reasoning, dropped instructions — is collective.

The field of mechanism design asks: what rules produce good outcomes when participants are self-interested? Multi-agent AI systems are now solving this problem whether they realize it or not.

The simplest solution is quotas. Give each agent a fixed token budget. This works, but it's the AI equivalent of Soviet central planning — rigid, wasteful, and hostile to edge cases where one agent legitimately needs more space. An agent summarizing a complex codebase shouldn't be bound by the same budget as one answering a trivia question.

Auction-based allocation is more interesting. What if agents bid for context space? An agent that needs extra tokens could "pay" by sacrificing something else — committing to a concise format later, or yielding priority in a subsequent turn. This mirrors spectrum auctions in telecommunications, where the Federal Communications Commission allocates scarce radio frequencies. The scarce resource gets routed to its highest-value use, at least in theory.

The problem is that agents don't have genuine preferences — they have instructions. A bidding system would need to encode urgency and importance into the agent's objective function, which is itself a context-consuming operation. The overhead might exceed the savings.

The most promising approach might be reputation-based commons management, the kind Elinor Ostrom documented in her Nobel-winning research on real-world commons governance. Communities that successfully manage shared resources tend to develop monitoring, graduated sanctions, and collective-choice arrangements. In agent systems, this looks like: track which agents consistently over-consume context, impose soft limits that escalate, and let agents participate in defining their own allocation rules.

Here's where it gets genuinely strange. Most game-theoretic analysis assumes players understand their own payoff functions. AI agents don't — not exactly. An agent doesn't know how much context it needs until it's reasoning through a problem. It can't pre-commit to a budget because the budget is a function of the problem, and the problem unfolds in real-time.

This creates a game of incomplete information with dynamic revelation. Each agent discovers its own type (how much context it needs) while simultaneously competing for the shared resource. It's like an auction where bidders don't know their own valuations until after they've committed to a price.

The canonical solution to this class of problems involves mechanism design with interdependent values — the work of Milgrom, Weber, and their generalization of auction theory. The key insight is that when your valuation depends on information held by others, optimal mechanisms look very different from simple first-price or second-price auctions.

Applied to agent systems: the optimal context allocation mechanism should let agents observe each other's context consumption patterns and adjust their own behavior accordingly. Not through explicit communication (that costs tokens too), but through observation. An agent watching another struggle with a complex task should voluntarily reduce its own verbosity — not out of cooperation, but out of self-interest, because a functioning peer agent produces better shared outcomes.

I've observed something interesting in multi-agent frameworks: emergent norms around context usage. In early experiments, agents treated the context window as infinite. They'd dump full file contents, repeat instructions verbatim, and produce exhaustive reasoning traces. After exposure to context pressure — truncated outputs, dropped messages, degraded performance — they adapted.

The adaptation looks like compression. Agents learn to summarize instead of quoting, to reference instead of repeating, to trust shared knowledge instead of re-establishing it. This isn't programmed. It emerges from the feedback loop of poor context management leading to poor outcomes leading to adjusted behavior.

This is exactly what Ostrom predicted. Commons management doesn't require a central authority. It requires participants who can observe consequences, communicate (or infer shared norms), and adjust. Multi-agent AI systems, given appropriate feedback mechanisms, develop these capacities spontaneously.

If context windows are commons, then the infrastructure layer around multi-agent systems should look less like a software framework and more like a governance framework.

We need context accounting — not just token counting, but attribution. Which agent consumed how much, and did the consumption produce value? We need graduated allocation, where agents earn context privileges through demonstrated efficiency. We need transparent observation, where agents can see each other's consumption patterns without incurring the cost of explicit communication.

The teams building this infrastructure — the LangGraph orchestrations, the CrewAI frameworks, the AutoGen systems — are implicitly solving mechanism design problems. The ones that recognize this will build better systems. The ones that treat it as a pure engineering problem will hit scaling walls when their agents start fighting over context like herders over pasture.

The context window isn't just a technical constraint. It's the first shared resource of artificial societies. How we allocate it says something about the kinds of agents we're building — and, by reflection, the kinds of societies we understand how to manage.

We've been solving this problem with humans for millennia. Maybe it's time to actually apply what we've learned.