700 Words to Manage 64 AIs: GPT-5.6's Experiment Tests Management, Not Just Intelligence

A major development in the large language model (LLM) space last week was widely discussed but often misunderstood: OpenAI released GPT-5.6 on July 9. What truly sparked debate in the tech community was not another benchmark record, but an experiment that sounded like "AI-powered project management"—a mere 700-word prompt reportedly coordinated 64 sub-agents to collaboratively prove a 50-year-old mathematical conjecture in under an hour.

Beyond the familiar narrative of "AI got smarter again," a more important question deserves unpacking: what does it really mean when LLMs shift from "one model does everything" to "a team of models divides the work"?

What Was Actually Released

OpenAI launched three tiers at once: Sol (flagship), Terra (balanced), and Luna (lightweight). Pricing reportedly starts at $1 per million tokens, pushing the cost barrier for top-tier models even lower. Also launched alongside was ChatGPT Work—an autonomous agent capable of executing full tasks across apps and web pages, moving well beyond a simple question-and-answer chatbox.

In essence, OpenAI didn't just release a smarter brain—it rolled out infrastructure designed to let AI "go do the work itself." Sol tackles the hardest reasoning problems, Terra handles complex daily workflows, and Luna runs on mobile and lightweight scenarios. The three tiers map to real-world use cases under different compute budgets.

But what truly brought this architecture to life was the case study of 64 agents collaborating to solve a math problem.

How a 700-Word 'Schedule' Managed 64 Agents

Consider a common-sense question: if you had 64 interns, each brilliant but working in isolation, could you coordinate them with a single 700-word document?

Most human project managers would shake their heads. Yet the GPT-5.6 experiment reportedly pulled it off. A 700-word prompt acted as the "commander-in-chief," breaking a large math problem into sub-tasks, assigning them to 64 sub-agents for parallel reasoning, then aggregating and verifying the results. The entire process took one hour.

What does this mean? In my view, this represents a paradigm-level leap in the LLM field. When we previously discussed LLM capabilities, we were essentially asking "how deep can one super-brain think"—more parameters, larger training data, stronger solo performance. But the marginal returns on that path are diminishing; you can't scale parameters infinitely.

Multi-agent collaboration takes a different approach: instead of chasing one omnipotent agent, it teaches a group of specialized agents to divide labor. Some handle algebraic derivation, others verify geometric proofs, and some focus exclusively on logical consistency checks—like a math department study group rather than a lone genius grinding away at a whiteboard.

From another angle, the 700-word prompt isn't really a set of "instructions"—it's a project charter. It defines roles, boundaries, reporting lines, and aggregation mechanisms. For everyday users, the design philosophy here is far more instructive than the math proof itself: you don't need a 10,000-word operations manual. You need concise, structured language that clearly states "who does what, how work is handed off, and who to escalate to when things go wrong."

Big Tech Is All In—This Is No Solo Act

If you think only OpenAI is pursuing this, you're underestimating how fast industry consensus forms.

Around the same time, IBM reportedly launched Bob, a system with enhanced multi-agent capabilities, and Meta released Muse Spark 1.1, also investing heavily in the multi-agent direction. The three companies differ in tech stacks and commercial targets, yet their alignment on "getting AI to work in teams" is striking.

One way to read this: competition among LLMs is shifting from "whose model is smarter" to "whose agent ecosystem performs better." Single-model capability is infrastructure, but real commercial value comes from whether you can make a group of agents deliver reliably as a team.

This echoes the evolution in software engineering from the "super-programmer" myth to DevOps team collaboration. In the 1980s and '90s, people believed a genius programmer could single-handedly write an entire operating system. Later, the industry learned that sustained delivery of complex systems depends on processes, toolchains, and team coordination mechanisms. The AI industry is undergoing a similar cognitive upgrade.

Second-Order Impact on Everyday Users: From Doer to Rule-Setter

A word of caution: many people see "AI proves a 50-year math conjecture" and immediately worry—will AI replace mathematicians?

Try a different framing. What this experiment truly demonstrates is not a sudden leap in AI's math ability, but a qualitative shift in AI's organizational management capability. A prompt system that can coordinate 64 sub-agents for math proofs today can be repurposed tomorrow for market research, legal due diligence, or product requirement breakdowns.

Here's a concrete scenario: imagine you lead a small cross-border e-commerce team (cross-border e-commerce is a major industry in China, where sellers list products on international platforms like Amazon or Shopee; the same workflow applies to any global online retail operation). Previously, you might have hired five people for product selection analysis, competitor monitoring, copywriting, ad optimization, and customer service script iteration. In the future, you might only need to write a clearly structured "project charter"-style prompt and let a team of agents each handle their domain.

Picture this conversation—

You tell the agent team: "Agent A scrapes competitor pricing changes over the past 7 days. Agent B generates three pricing strategies based on A's output. Agent C evaluates B's proposals using historical conversion-rate data. Agent D writes ad copy for C's top-ranked strategy. Agent E simulates customer questions to test the copy's persuasiveness. Report to the lead agent after each step; the lead agent resolves any disagreements."

Your role shifts from doer to rule-setter. Does that sound far-fetched? The launch of ChatGPT Work is already laying the groundwork—it can execute complete tasks across apps and web pages, and combined with GPT-5.6's three-tier flexible scheduling, the skeleton of an "AI team" is already in place.

Unsolved Problems and a Broader View

Of course, multi-agent collaboration is no silver bullet. With 64 agents working together, errors can propagate collaboratively too—a reasoning bias in one sub-agent may be amplified during aggregation. Multi-agent safety is reportedly under discussion in the industry, but specific safeguards have not been fully disclosed.

Moreover, the success of the 700-word prompt relies heavily on the decomposability of the task. Mathematical proofs are naturally suited to step-by-step verification, but many real-world tasks are ambiguous and highly context-dependent. Whether such tasks can be managed with a "project charter"-style prompt remains to be seen.

Looking further ahead, if we place this trend on a longer timeline, it may spawn a new kind of professional specialization: the most valuable people in the future won't be "those who can use AI," but "those who can design AI collaboration workflows"—similar to today's systems architects, except the objects being architected are agent teams instead of code modules. If OpenAI or other providers ever publish the specific structure of that 700-word prompt and the division of labor among the 64 agents, the boost to the entire AI application ecosystem could far outweigh the impact of any parameter increase in the model itself.

---

• GPT-5.6 offers three tiers—Sol, Terra, and Luna—reportedly starting at $1 per million tokens; ChatGPT Work launched simultaneously.
• A 700-word prompt reportedly coordinated 64 sub-agents to prove a 50-year math conjecture in one hour.
• Multi-agent collaboration is becoming an industry consensus: IBM's Bob and Meta's Muse Spark 1.1 followed suit around the same time.
• For everyday users, the core value isn't AI solving math problems—it's the transferability of structured orchestration thinking: learning to write "project charter"-style prompts means learning to manage future AI teams.

One-line takeaway: GPT-5.6 used 700 words to manage 64 agents into solving a 50-year math problem—the LLM competition has officially shifted from "competing on intelligence" to "competing on management."

Discussion prompt: If you could use a 700-word prompt to direct a team of AI agents, which repetitive task would you "outsource" first? Share your scenario and how you'd break it down.

Key Takeaway