複数のAIに足りないのは、賢さではなく共有コンテキストだったThis article argues that the real bottleneck in multi-agent AI systems is not…

agent">Multi-agent AI systems, in which several language-model-driven agents divide a task and collaborate, have become one of the more heavily promoted architectural patterns of the past two years. This article makes a pointed argument about why many such systems disappoint in practice: the limiting factor is rarely the intelligence of any single agent, but the absence of shared context between them. If accurate, that reframing matters because it shifts engineering effort away from chasing more capable models and toward designing the information substrate that agents operate on together.

The core claim is straightforward. As individual models have grown more capable, the marginal gains from swapping in a stronger model have narrowed for many coordination-heavy tasks. What breaks down instead is the handoff. When one agent completes a subtask and passes work to another, much of the reasoning, intermediate state, and implicit assumptions behind that work do not travel with it. Each agent tends to operate within its own context window, and language models are fundamentally stateless between calls, so anything not explicitly encoded in the prompt is effectively forgotten. The result resembles a game of telephone, where information degrades at every boundary even if each participant is individually competent.

This framing helps explain a common failure mode. A planner agent may produce a sound decomposition, a researcher agent may gather correct facts, and a writer agent may be fluent, yet the final output is inconsistent, repeats work, or contradicts earlier decisions. The article attributes this not to any agent being unintelligent, but to the fact that no agent has a complete, current view of what the others know and have decided. Individual competence does not compose into collective competence without a mechanism to carry context across the seams.

The proposed remedy is to treat shared context as a first-class design concern rather than an afterthought. Several established patterns are relevant here. A blackboard architecture, a concept that predates the current wave of AI, gives agents a common workspace they can read from and write to, so state lives in one place rather than being copied through messages. Shared memory stores, often backed by vector databases for semantic retrieval, let agents pull relevant history on demand instead of relying on whatever fits in a single prompt. Structured message passing, where handoffs include explicit summaries of goals, constraints, and prior conclusions, reduces the loss that occurs during transfers. Each approach trades off differently between completeness, token cost, and latency, and the article's practical point is that these trade-offs deserve deliberate attention.

This discussion sits within a broader industry effort to standardize how models access and exchange context. Anthropic's Model Context Protocol, introduced in late 2024, aims to provide a common interface between models and external data sources and tools, which is adjacent to the shared-context problem even though it targets tool and data connectivity more than inter-agent state. Orchestration frameworks such as LangGraph, Microsoft's AutoGen, and CrewAI each offer their own conventions for passing state between agents, and their design choices reflect the same underlying tension. Retrieval-augmented generation, now a mainstream technique, is essentially context injection for a single model, and agent">multi-agent context sharing can be seen as a generalization of that idea across multiple actors.

Some caveats are worth keeping in mind. The claim that context, rather than capability, is the dominant bottleneck is likely task-dependent; for problems that hinge on hard reasoning within a single step, a stronger model still helps. Sharing more context also carries costs, since larger prompts increase token usage, raise latency, and can introduce noise that degrades output, so more shared context is not automatically better. The article appears to argue for well-designed context sharing rather than maximal context.

For practitioners, the takeaway is a shift in emphasis. Before reaching for a more powerful model or adding another agent, it is worth examining what information is lost at each handoff and whether a shared memory, a common workspace, or richer structured messages would recover it. Framed this way, building effective agent">multi-agent systems looks less like assembling clever specialists and more like engineering the shared environment in which they think together.