What This Is
OpenAI published a counterintuitive dataset this week: GPT-5.6's ultra mode defaults to running 4 agents — AI modules capable of independently completing subtasks — in parallel on the same task. The result: higher scores, shorter runtimes, and lower costs. This means the decisive factor in AI is shifting from "the model itself" to "system orchestration." NVIDIA's AVO also used a multi-agent architecture to score a perfect result on the ARC-AGI-3 abstract reasoning benchmark.
Multi-agent targets the structural weakness of a single agent: one large model crams planning, execution, and self-check into the same context window (the amount of text the model can "see" at once). The longer the context, the more likely the model forgets constraints set in step one by step six — at root, a pipe-width problem. Running 4 agents in parallel splits that pipe into four.
The engineering puzzle has three pieces: task decomposition (the orchestrator splits hard problems among planners, executors, critics, and verifiers), communication isolation (agents pass information via a shared blackboard or peer-to-peer messages — the core principle is no contamination), and result aggregation (using self-consistency voting, Best-of-N, and similar strategies to synthesize the final answer).
Industry View
Supporters view this as an inevitable direction. OpenAI, NVIDIA, and Anthropic have released multi-agent frameworks in dense bursts over the past three months — evidence that the "bigger model equals stronger model" path is ceding ground to the engineering dividend of "system orchestration."
But skepticism is equally explicit. An MIT study shows multi-agent systems have higher failure rates than single agents on complex tasks: communication overhead and coordination errors scale exponentially. Practitioners also point out that the "cheaper" conclusion rests on specific benchmarks. Switch to real business data — contract review, for instance — and subtask boundaries aren't as clean; the cost advantage can vanish in an instant.
Another underweighted risk is accountability: when a report produced jointly by 4 agents contains errors, who takes the blame? No technical solution exists today.
Impact on Regular People
For enterprise IT: procurement lists must change. Multi-agent deployment demands orchestration frameworks, communication monitoring, fault tolerance — budgets should shift from "model" to "system."
For individual careers: AI tools are shifting from "a single dialog box" to "a collaborating team." In the next year or two, "explaining a problem clearly" will be more valuable than "asking cleverly."
For the consumer market: regular users can't yet touch 4-agent mode. It's still a hidden switch behind per-call pricing — only after API prices drop two more rounds is it likely to reach personal subscriptions.