Multi-agent AI is pitched as a fix for memory limits, but does it just move the problem?

This explores whether the limits on what a single agent can hold and manage in its context are the same problem as the difficulty of getting several agents to coordinate. The corpus suggests they are two views of one problem: deciding what information reaches which agent, when, and whether anyone checks it. Multi-agent systems are often pitched as an escape from context limits, but they mostly move the problem somewhere else.

The case for splitting work across agents is real. Tasks that need different kinds of expertise, parallel execution, and independent verification exceed what one agent loop can organize, and more capability or a bigger window doesn't fix that Do single agents always hit organizational limits?. But one line of work shows that a lot of what looks like a multi-agent need is really a memory-management need. The Thread Inference Model structures reasoning as recursive subtask trees and prunes the KV cache by rule. It keeps reasoning accurately past the context limit, even while manipulating 90% of the cache, so a single model can do the job of a multi-agent system Can recursive subtask trees overcome context window limits?. Another finding points the same way: about 80% of multi-agent performance variance comes from token budget, not coordination cleverness How does test-time scaling work at the agent level?. Much of the apparent gain from many agents is simply more context to spend.

Once you do split the work, the context problem reappears between the agents. On the AgentsNet benchmark, coordination degrades predictably as the network grows. Agents agree too late, or adopt a strategy without telling their neighbors. They also accept neighbors' information without verifying it, so errors spread, even though the same agents can spot a direct conflict Why do multi-agent systems fail to coordinate at scale?. That is a gating failure: whatever lands in an agent's context is trusted by default. Single agents have the same flaw. Long-workflow failures come from weak memory control, because transcript replay and retrieval have no gate on what gets kept, not from missing knowledge Can agents fail from weak memory control rather than missing knowledge?. Agents chatting with each other amounts to transcript replay across a whole team. The deliberation failures the corpus names, Degeneration-of-Thought and Silent Agreement, fit that pattern What limits autonomous capability in large language models?.

The fixes point the same way: control the channel instead of enlarging it. MetaGPT-style agents hand each other standardized documents and pull what they need from a shared environment, not push conversation at each other, which cuts noise and beats chat-style coordination Does structured artifact sharing outperform conversational coordination?. The single-agent equivalent is a bounded, schema-governed committed state that separates temporary recall from permanent memory writes Can agents fail from weak memory control rather than missing knowledge?. Viewing agents as computational graphs makes the link explicit: nodes are operations, edges are information flow, and Chain-of-Thought, Tree-of-Thought, and Reflexion turn out to be the same kind of structure. Both the prompts at the nodes and the wiring between them can be optimized automatically Can we automatically optimize both prompts and agent coordination?. On this view, deciding who talks to whom is context engineering applied to a team.

So the useful question is less whether to use one agent or many, and more what the cheapest way is to make sure each step sees only what it needs and that anything it trusts has been checked. Sometimes a single model with disciplined memory answers that. Sometimes you need several agents, and then the design work goes into the channels between them.