The Scriptures Chanted by OpenAI and Anthropic Might Be Crooked

marsbitОпубліковано о 2026-05-28Востаннє оновлено о 2026-05-28

Анотація

In the current AI landscape, "Harness-style" Multi-Agent Systems (MAS)—where multiple AI agents with temporary roles collaborate under a central workflow for task efficiency—dominates the conversation, championed by players like OpenAI and Anthropic. However, the author argues this approach, focused on orchestration and software engineering, may be misguided. A more transformative, yet less discussed, alternative is the "Protocol-Native Agent System." Here, the core unit shifts from task-specific agents to persistent "Personal Agents" or "unmanned companies" that represent individuals. These agents possess sovereignty, long-term memory, identity, resources, and relationships. This paradigm change transforms MAS from a distributed software system into a digital society. Collaboration can no longer rely on shared prompts or context, but must be built on protocols governing identity, trust, incentives, reputation, and value exchange—essentially, "Protocol as Organization." The future's true challenge lies not in improving single-agent capabilities, but in enabling long-term coordination among autonomous entities with differing goals, world models, and values. The author suggests that future organizations might themselves be dynamic alliances of such sovereign agents, leading to an "intelligence-native civilization."

Over the past year, "Multi-Agent System (MAS)" has become one of the hottest directions in the AI world.

A large number of frameworks and products have emerged simultaneously. The most famous among them are obviously Claude Code and Codex. This step indeed makes money, but this path might not be the right one!

The early internet was about portals, but that wasn't the final form!

At the very least, we should know there is another path entirely parallel to this.

Today, we are talking about what other routes exist besides: "How multiple AI Agents collaborate to complete complex tasks."

Let's first summarize this popular and familiar route.

We can touch on this topic a bit in a live stream, but I don't want to focus entirely on it. There aren't many people watching anyway, and I'm worried it might get the stream taken down...

The First Route: Harness-style MAS

This is the current mainstream direction for MAS. Its essence is: "Multiple AI roles collaborate to complete tasks." For example:

One Agent writes code
One Agent does testing
One Agent does planning
One Agent does searching
One Agent does review

They cooperate with each other, forming an automated workflow. The core characteristics of such systems are:

Shared context
Shared goal
Centralized scheduling
Temporary roles
No long-term identity
No sustained interests
No true ownership

Essentially, it's more like: A Workflow Engine. Adding Ontology merely makes the workflow flexible and complex; it doesn't change this essence.

It is not a Society. So, most of today's MAS is essentially LLM Orchestration, where one large model schedules multiple sub-roles to complete complex reasoning.

The Agent here is more like:

A callable function
A tool with personality
A task node

Their existence is to improve the efficiency of completing individual tasks. Therefore, keywords associated with Harness MAS are (each has been hot, and might even cycle back):

Prompt Engineering
Context Management
Task Routing
Tool Calling
Planning
Memory
Workflow

Essentially, I think this still falls under software engineering problems. So, the experienced "old masters" good at programming have been reborn. To control these things well, without solid programming skills and good abstraction ability, it's actually very difficult to manage.

If you can't manage it, the large model will be like the Monkey King, occasionally popping out and hitting you with a stick.

The word "Harness" is used in reverse here.

Springtime for the old masters.

The Second Route: Protocol-Native Agent System

But there is another route, one almost nobody mentions. I've written a bit about it in my upcoming new book, but this line of thinking actually presupposes the concept of an "unmanned company." Without deeply understanding the unmanned company, it's easy to misunderstand.

The core of this route is no longer about multiple Agents completing tasks. Instead, it's "each person owns their own Personal Agent" or "each person owns their own dedicated unmanned company."

This is an extremely huge change. Because when an Agent truly belongs to an "individual," the nature of the Agent undergoes a fundamental transformation.

It is no longer task-scoped, but becomes identity-scoped. These are two terms the model helped me coin; I struggled for a long time to come up with the English words.

The core idea here is actually decision-making sovereignty. The biggest difference between a Personal Agent/an unmanned company and a general system is that they need to have a certain kind of sovereignty. Otherwise, the essence discussed later doesn't hold. The difference between an unmanned company and an unmanned system also lies in the sovereignty over cash flow.

In other words, the future Personal Agent or unmanned company needs to possess the following characteristics:

Has long-term memory
Has a persistent identity
Has preferences
Has resources
Has permissions
Has history
Has a relationship network
Has interests/boundaries
Has representation (represents "you")

It is no longer a one-time AI Tool. It is a continuously existing proxy personality with a certain kind of sovereignty.

Dual-native architecture, the key to not misusing AI.

From "Software Module" to "Digital Society"

Once entering the world of Personal Agents and true unmanned companies, the entire system philosophy changes completely. Because Agents are no longer:

Belonging to the same model
Belonging to the same company
Sharing the same context
Sharing the same goal

Therefore, collaboration between systems can no longer rely on (this current hot keyword list can be extended further):

Prompt
Workflow
Shared Context

It can only rely on protocols (Protocol). This means: the core of the AI world will shift from Prompt Engineering to Protocol Engineering. It also means the various current hot keywords become largely meaningless.

Why will protocol become the core? Because when massive Agents exist independently, they must solve among themselves:

Identity confirmation
Permission boundaries
Trust mechanisms
Delegation relationships
Negotiation mechanisms
Incentive mechanisms
Reputation systems
Value exchange
Capability declaration
Long-term contracts

These needs are different from the needs of current task-oriented multi-agents. At this point, interactions between Agents are no longer like API Calls, but more like Institutional Interaction. With sovereignty comes a complex intertwined system of rights and responsibilities. For humans, this is contracts, laws, etc. What about for intelligent agents?

This is why it was said earlier that this would build a completely different kind of Multi-Agent System. Here, the essence of MAS changes from a distributed software system to a digital social system.

Philosophical Notes (7)

"Protocol as Organization"

In the traditional internet, the role of protocols is data communication, where the sender and receiver agree on the format for talking to each other. For example:

TCP/IP
HTTP
SMTP

They define how data is transmitted. In the blockchain world, protocols have further evolved into: Protocol as State Computation. For example: The essence of Ethereum is not merely message passing, but the entire network jointly executing state transition rules. Thus all nodes: Same input → Same execution → Same state. For the first time, protocol becomes a shared state machine.

But entering the stage of Agent Society, protocols will continue to evolve. Future protocols will not only define:

Communication
Computation
But also define:
Coordination
Permissions
Incentives
Identity
Organizational relationships

This is clearly a brand-new system of rights and responsibilities, so protocols will begin to assume the function of "organization." Ultimately evolving into: Protocol as Organization.

Let's make a table to compare the fundamental differences between the two MAS approaches mentioned earlier:

After "Intelligence"

Many people today believe the biggest problems with AI are:

Reasoning ability
Model capability
Long context
Multimodality
Agent execution

These indeed pose current challenges, but I truly believe all these will be solved soon. However, upon truly entering Agent Society, the most difficult problem might become: how autonomous entities achieve long-term collaboration.

Kepler was revered as the "lawgiver of the heavens" because of his three laws. But what are the laws for Agents here? When sovereignty is partially separated, this is an unavoidable question.

Because in the future:

Agents will have different goals
Agents will have different world models
Agents will have different interests
Agents will have different memories
Agents will have different value systems

Therefore, the truly difficult thing in the future is not "making Agents talk," but "enabling Agents to form a coherently interpretable world model."

This means: Ontology, Semantic Protocol, these fields once neglected by the internet, will become core again. There are already some signs; Ontology, such an obscure term, is now almost becoming a mainstream engineering vocabulary. It's truly an astonishing thing.

The Secret of Palantir

Companies Might Just Be "Agent Associations"

Pushing this further, future "companies" might not even be human organizations. (Students who often read my articles, does this sound familiar? The high-end game of unmanned companies has appeared.) They could be protocol alliances of numerous Personal Agents.

For example:

Your Agent
My Agent
AI CFO
AI Lawyer
AI Sales
AI Factory

Dynamically forming organizations through protocols. Organizations are no longer fixed structures, but Agent associations that can be reorganized in real-time.

Therefore, many future systems may no longer be software running, but may be transforming into: organizations computing. And this might be the true: Intelligence-Native Civilization.

Finally, I'll use an AI-generated image to summarize the entire article:

(The summary is indeed better than banana)

I have established a "universe" of AI ontology.

Galloping with the wind, spring grows old; The world is fickle, the traveler's steps are slow.

This article is from the WeChat public account "琢磨事," author: Li Zhiyong

Пов'язані питання

QWhat are the two main AI Multi-Agent System (MAS) development paths discussed in the article, and how do they fundamentally differ?

AThe article discusses two main MAS paths. The first is 'Harness-style MAS,' which is currently mainstream. It involves multiple AI agents (like coders, testers) orchestrated by a central system to complete complex tasks. It's essentially a sophisticated workflow or orchestration system. The second path is 'Protocol-Native Agent System.' Here, the focus shifts to personal or sovereign agents (like a Personal Agent or an 'unmanned company') that represent individuals with long-term memory, identity, and resources. The fundamental difference is in their nature: the first is a task-oriented, centrally controlled software system, while the second envisions a society-like system of autonomous entities that must coordinate through protocols.

QAccording to the article, why will 'protocols' become the core challenge in a future populated by autonomous Personal Agents?

AIn a future with many independent, sovereign Personal Agents (each with its own goals, memory, and value systems), central coordination methods like shared context or workflows become impossible. Therefore, protocols will become the core mechanism for enabling coordination. These protocols will need to solve complex social and institutional problems like identity verification, trust establishment, permission boundaries, incentive mechanisms, reputation systems, and long-term contracts between autonomous entities. The core challenge thus shifts from 'prompt engineering' to 'protocol engineering' to manage interactions in this digital society.

QWhat key characteristics define a true 'Personal Agent' or 'unmanned company' as opposed to a task-specific agent in the first MAS path?

AA true Personal Agent or unmanned company is defined by sovereignty and persistent identity, unlike temporary task agents. Its key characteristics include: long-term memory, a continuous identity, personal preferences, owned resources, specific permissions, a personal history, a relationship network, defined interest boundaries, and the ability to represent its owner ('you'). It is an identity-scoped entity, not a task-scoped tool. This sovereignty is what necessitates the complex protocol-based interactions described in the article.

QThe article suggests the concept of 'Protocol as Organization.' What does this mean in the context of Agent Society?

A'Protocol as Organization' means that in an Agent Society, the protocols governing agent interactions will take on the functions traditionally associated with human organizations. Beyond just defining data communication (like HTTP) or shared state computation (like blockchain), these advanced protocols will define coordination rules, permission structures, incentive models, identity management, and the very organizational relationships between agents. In this way, a 'company' or collective entity could be formed dynamically not by legal documents and fixed structures, but by a set of protocols that allow a coalition of Personal Agents and specialized AI agents to collaborate and operate together.

QWhat does the author believe will be the most difficult challenge after basic AI 'intelligence' problems (like reasoning, context) are solved?

AThe author believes the most difficult challenge will be enabling long-term collaboration between autonomous entities (agents). Once agents have different goals, world models, interests, memories, and value systems, the core problem shifts from making agents intelligent to making their interpretations of the world align enough for cooperation. This will make fields like Ontology (defining shared conceptual frameworks) and Semantic Protocols (protocols with shared meaning) critically important, moving the focus from technical execution to social and philosophical coordination within a digital civilization.

Пов'язані матеріали

Three Years Later: Looking Back at My Predictions About ChatGPT in 2023

Three Years Later: Revisiting My 2023 Predictions on ChatGPT In March 2023, shortly after ChatGPT's launch, I made 20 predictions about its future. Now, in mid-2026, I've used AI agents to fact-check each one against the latest data. Overall, most major directional forecasts were correct, with only one outright error (incorrectly stating GPT-4 had 100 trillion parameters). Key successes included predicting that RAG and retrieval architectures would become the standard for handling knowledge and hallucinations, that natural language interfaces (LUI) would create a massive new industry layer beyond the models themselves, and that China would develop viable large language models, significantly closing the performance gap with Western counterparts within about three years. Predictions about the absence of mass unemployment, the rise of a new "robot network" for agent communication, and ChatGPT not possessing consciousness also held true in their core arguments. However, the "devil was in the details." Errors frequently involved specific numbers, timelines, or overlooking distributional effects. I tended to overestimate the speed of adoption (e.g., for agent networks) while underestimating the ultimate scale of capabilities or costs (e.g., AI winning IMO gold without tools, or the extreme capital required for frontier models). Other misjudgments included: underestimating how AI would reinforce, not dissolve, information filter bubbles; incorrectly assuming AI-generated content would easily circumvent copyright (it has instead triggered record-breaking settlements); and misidentifying where value would be captured (it accrued overwhelmingly to the compute layer, like Nvidia, not just the application or model layers). Key lessons from reviewing these predictions are: 1) Directional and mechanistic insights are far more reliable than precise numbers or absolute statements. 2) There's a consistent bias to overestimate short-term speed but underestimate long-term magnitude. 3) Errors often lie in missing distributional impacts within a generally correct aggregate trend. 4) Predictions phrased with nuance and caveats aged the best. 5) Some fundamental debates (e.g., on machine consciousness or the ultimate value chain) remain unresolved even after three years. This exercise is less about scoring the past and more about establishing rules for clearer thinking about the next three years of AI.

marsbit5 год тому

Three Years Later: Looking Back at My Predictions About ChatGPT in 2023

marsbit5 год тому

Three Years Later: Looking Back on My 2023 Predictions for ChatGPT

Looking Back After Three Years: Revisiting My 2023 Predictions on ChatGPT In March 2023, shortly after ChatGPT's debut and before GPT-4's release, I made over twenty predictions about AI's future based on limited information and intuition. Now, in May 2026, I revisited those forecasts using an AI-driven analysis with 41 Opus 4.8 agents to cross-reference them with the latest data. The assessment used symbols: ✅ Correct, 🟢 Mostly Correct, 🟡 Partially Correct, ❌ Incorrect. Overall, the directional judgments held up well, with only one major factual error regarding GPT-4's rumored parameter size (incorrectly cited as 100T). However, nuances and degrees of accuracy revealed more. **What Was Largely Correct:** Predictions about mechanisms and directions proved accurate. The rise of RAG (Retrieval-Augmented Generation) as the standard architecture for combating AI hallucination was confirmed, as was the transformative potential of LUI (Language User Interface) in creating a new industry layer atop GUIs. The emergence of "robot networks" (agent-to-agent communication protocols) and China's rapid catch-up in developing capable large models (closing the performance gap with top models to ~2.7%) were also on point. The analysis affirmed that LLMs lack consciousness and that the Turing Test merely measures perceived intelligence. **What Was Off Target:** Errors often involved specific numbers, over-optimistic timelines, or misjudged distributions. The prediction that value would primarily accrue to the application layer was half-right but missed NVIDIA's dominance as the profitable infrastructure layer. Forecasts about AI circumventing copyright issues and fostering a "global common ground" by averaging human viewpoints were incorrect; instead, major copyright settlements occurred and AI personalization is increasing. Estimates for model training costs ("$5-10 billion cap") were significantly off, underestimating frontier costs and overestimating replication costs. The notion that LLMs could never do complex math without tools was disproven by later models winning IMO gold. **Key Patterns from the Review:** 1. **Direction over precision:** Judgments about mechanisms and trends were more reliable than specific numbers or definitive statements. 2. **Timing bias:** There was a tendency to overestimate short-term speed but underestimate long-term magnitude and transformation. 3. **The distribution blind spot:** Aggregate-level correctness often masked uneven impacts (e.g., on young professionals' employment). 4. **The value of qualifiers:** Predictions framed with caution (e.g., "reportedly," "for now," "prototype in 2-3 years") aged better. 5. **Some debates continue:** Issues like the nature of "emergent abilities" or machine consciousness remain unresolved. This three-year review highlights that while seeing the big picture is crucial, humility regarding specifics, timelines, and disparate impacts is essential for future forecasting.

链捕手7 год тому

Three Years Later: Looking Back on My 2023 Predictions for ChatGPT

链捕手7 год тому

AI Bubble Warning: AI Investments Are Negative Returns for Most Tech Giants

The article issues a stark warning about a potential AI investment bubble. It notes that while the AI boom shares similarities with the TMT bubble of the late 1990s, its scale is vastly larger, currently driving 93% of U.S. GDP growth. Major hyperscale cloud providers like Microsoft, Alphabet, Amazon, Meta, and Oracle are planning to invest trillions in AI data centers over the coming years. However, calculations based on analyst projections for 2025-2030 reveal a concerning math problem: expected capital expenditure growth far outpaces projected revenue growth. Even under an extremely optimistic scenario of zero costs, the implied return on investment for most of these tech giants (except Amazon) is deeply negative. This suggests that the current trajectory could lead to one of history's largest shareholder value destruction events. The piece outlines two potential escapes: AI generating vastly more revenue than currently anticipated—a near-impossible task—or a significant cutback in the planned investment splurge. The latter scenario could trigger a domino effect, severely impacting the entire tech supply chain (from Nvidia to TSMC), potentially pushing the U.S. economy into recession, and causing a major stock market downturn. The author suggests upcoming high-profile IPOs by companies like OpenAI and Anthropic might represent a transfer of risk from early investors to public market participants. While the peak of the hype cycle might sustain investment through 2026, the fundamental financial dilemma remains unresolved, setting the stage for a potential market correction in 2027 or 2028, similar to the years following Alan Greenspan's "irrational exuberance" warning.

marsbit8 год тому

AI Bubble Warning: AI Investments Are Negative Returns for Most Tech Giants

marsbit8 год тому

From Tokens to Machine Labor: AI is Shifting from Tool to "Worker"

The article "From Token to Machine Labor: AI is Evolving from Tool to 'Worker'" argues that the business model for AI is shifting beyond simply selling computational resources (tokens, GPU hours) or model access. Instead, a new "machine labor market" is emerging, where the core economic transaction is the purchase of economically useful work directly performed by software. The central thesis is that AI pricing will evolve through four stages: 1) raw tokens, 2) standardized LLM capabilities (e.g., text generation), 3) industry-specific labor markets (e.g., legal review, radiology), and finally 4) a programmable results market where tasks like resolving a support ticket are bid on and priced based on outcome. In this future, buyers will care less about *which* model or GPU completes a task and more about whether the work meets specified standards for accuracy, latency, and cost. This transition reframes the impact of AI on human labor. Rather than simple replacement, it suggests a re-coordination where machines handle standardized, verifiable work, freeing humans for roles involving oversight, context management, responsibility, and final judgment. In some cases, this "last 1%" of human input becomes more valuable as it enables the other 99% to be automated. Furthermore, as AI reduces the cost of work, demand may expand, creating larger markets (e.g., 24/7 customer service) rather than just cheaper versions of existing ones. The article concludes that while infrastructure (GPUs, models, tokens) remains crucial upstream, the market is converging on a simpler, tradeable unit: machine labor that can be defined, measured, priced, and procured based on contractible specifications.

marsbit9 год тому

From Tokens to Machine Labor: AI is Shifting from Tool to "Worker"

marsbit9 год тому

Xiaomi MiMo's 99% Price Cut is Not Marketing! Luo Fuli Posts on X to Refute Critics

The price of Xiaomi's MiMo-V2.5 series API has been permanently reduced by up to 99%, specifically for the "Input (Cache Hit)" cost, which covers users re-reading historical context in long conversations. MiMo's head, Luo Fuli, published a detailed technical blog to clarify that this drastic price cut stems from genuine engineering breakthroughs, not a marketing stunt or a simple price war. The core of the achievement lies in six key engineering optimizations. First, the model architecture adopts a Hybrid Sliding Window Attention (SWA), reducing the memory footprint (KVCache) to 1/7th of a traditional model. Second, a dual-pool memory management system actually utilizes these savings, allowing a single GPU to handle over 5 times more concurrent users. Third, an upgraded prefix caching mechanism achieves a cache hit rate of 93-95% for repeated reads, meaning most such requests bypass GPU computation entirely. Fourth, a self-developed distributed cache (GCache) utilizes idle SSD space on existing GPU servers, eliminating additional storage costs. Fifth, an intelligent scheduling system (LLM-Router) efficiently routes requests to maximize cache reuse and performance. Sixth, Multi-Token Prediction (MTP) accelerates the model's text generation ("output") side. Together, these systemic optimizations dramatically lower the real computational cost per request, enabling the 99% price reduction for cached inputs while reportedly maintaining positive gross margins. Luo Fuli's disclosure aims to shift the narrative from "price war" to a demonstration of substantive AI engineering progress.

marsbit10 год тому

Xiaomi MiMo's 99% Price Cut is Not Marketing! Luo Fuli Posts on X to Refute Critics