← All articles

ARTICLE · TODD KELSEY

Agent Wars III: MANGO-X Meets D-BAM's Qwikido

On LinkedIn ↗

中文

A fictional roundtable for the real contest between American and Chinese frontier AI labs

Put the leading American and Chinese AI labs in one room and the first surprise is that they are not fighting the same war.

This is a thought experiment, not a report of an actual summit. Around the table sit six American companies I will call MANGO-X: Microsoft, Anthropic, NVIDIA, Google, OpenAI and xAI. Their five public-facing agent teams form the Fab Five: Copilot, Claude, Gemini, ChatGPT and Grok. NVIDIA is the sixth member because every superhero team eventually discovers that its armorer, foundry and power plant belong in the story.

Across from them sits D-BAM: DeepSeek, ByteDance, Alibaba and Moonshot AI. Their agent team is Qwikido: DeepSeek, Doubao, Qwen and Kimi. A fifth chair goes to MiniMax, whose multimodal reach makes it too important to leave outside merely because it spoils the acronym.

The names are playful. The structural difference is not.

Two industrial ecologies walk into a room

MANGO-X arrives as a vertically layered system.

  • Microsoft owns an installed enterprise estate and a distribution surface measured in organizations rather than model downloads.
  • Anthropic concentrates on reliable, deep work across research, writing and code.
  • NVIDIA supplies the machines, software and economics beneath nearly everyone else's ambition.
  • Google combines frontier research, cloud infrastructure, distribution and one of the world's largest knowledge systems.
  • OpenAI pushes toward a general-purpose human workbench that can reason, use tools and complete work across applications.
  • xAI brings real-time public conversation, rapid infrastructure expansion and an appetite for persistent agents.

D-BAM arrives through a different route.

  • DeepSeek turned efficiency and open deployment into strategic prestige.
  • ByteDance can move models into consumer products and media systems at extraordinary speed through Doubao and its broader ecosystem.
  • Alibaba's Qwen family joins open models to cloud distribution and a huge developer surface.
  • Moonshot's Kimi has made long context, research and agentic knowledge work central to its identity.
  • MiniMax occupies the fifth chair because frontier competition is increasingly multimodal: text, voice, image and video are converging into one creative and operational layer.

Calling these the “top five” is an editorial map, not a permanent league table. Tencent, Zhipu AI and other Chinese teams can reasonably claim a chair depending on whether the measure is research, distribution, open-model usage, capital, enterprise adoption or consumer reach. The important point is the shape of the field: China no longer has one conspicuous challenger. It has a dense competitive ecology.

The first question is not who has the smartest model

The chair asks each side a different question.

To MANGO-X: How much of the economy can your agents enter, understand and operate?

To D-BAM: How much intelligence can you make cheap, adaptable, distributable and unavoidable?

Those are not mirror images. The American advantage is still concentrated in frontier compute, capital, cloud control planes, enterprise relationships and premium integrated products. The Chinese advantage is increasingly visible in efficiency pressure, open-model availability, ruthless product iteration and the ability to join models to enormous domestic platforms.

The first Agent Wars could be imagined as a race toward a locked door. The second became a workbench for one person. The third is a meeting because the agents are becoming infrastructure. Once models can call tools, retain context, write and execute code, operate browsers, generate media and coordinate other agents, the contest shifts from a benchmark leaderboard to an industrial question: whose systems become the default labor layer?

MANGO-X discovers that NVIDIA is not merely a supplier

At the American end of the table, the Fab Five speak in product names: Copilot, Claude, Gemini, ChatGPT and Grok. NVIDIA says little. Its machines hum underneath the room.

That silence is the point. NVIDIA is not the sixth chatbot. It is the industrial substrate that helps determine the price, speed and geopolitical location of frontier intelligence. Microsoft and Google add cloud capacity. OpenAI, Anthropic and xAI turn compute into behavior. The public sees five agents; the system depends on a sixth participant that manufactures the possibility space.

This creates strength and fragility at the same time. The American stack can deliver tightly integrated, high-capability systems, yet it is capital intensive and concentrated. Export controls, power availability, chip supply, data-center construction and cloud bargaining power become part of model strategy.

D-BAM discovers that open models are distribution

At the Chinese end, “open” does not mean one thing. Licenses vary. Training data remain obscure. Product layers may be closed even when model weights are available. Still, open-weight distribution changes the economics of agent adoption.

An enterprise, developer or country that cannot justify a premium closed model for every step of a long agentic workflow can route more work through efficient open models. Agent systems consume many tokens while planning, calling tools, checking results and trying again. In that world, a model that is slightly less capable but dramatically cheaper—or deployable under local control—may win far more work than a leaderboard suggests.

This is why Qwen, Kimi, DeepSeek and MiniMax matter beyond China. They are not only national champions. They are components that can be inserted into global agent stacks. ByteDance adds a different form of power: the ability to connect intelligence to distribution, content and daily behavior at consumer scale.

Qwikido versus the Fab Five is the wrong matchup

The comic-book framing suggests a clean team battle. The real market will mix the rosters.

A U.S. company may run a Claude or OpenAI model for high-stakes reasoning, a Qwen or DeepSeek model for lower-cost subtasks, NVIDIA hardware underneath, Microsoft identity controls around the workflow and Google's data or search services feeding it. A Chinese developer may use open protocols and techniques created in the United States while optimizing for domestic chips and local platforms.

The Model Context Protocol captures part of this shift. When tools and data sources can expose a common interface, models become more substitutable and orchestration becomes more valuable. The durable advantage moves upward into memory, permissions, workflow design, evaluation and the ability to prove what an agent saw and did.

That is where Memory Atlas enters my own work. The more interchangeable the models become, the less sensible it is to let any one provider exclusively own an organization's working memory. The missing layer is portable continuity with provenance: not merely what an agent remembers, but which evidence it used, what authority it had and what state of the world existed when it acted.

What the fictional meeting should negotiate

No one at this table is going to agree on a universal theory of safety, speech, national security or data governance. A useful meeting would begin lower, with operational agreements that remain valuable even under rivalry.

  1. Incident vocabulary. Shared categories for agent failures involving unauthorized action, deceptive behavior, data leakage and loss of control.
  2. Evaluation transparency. Clearer disclosure of what benchmarks measure, how tool use is scored and where results depend on hidden scaffolding.
  3. Traceable action. Interoperable records of tool calls, source evidence, permissions and consequential changes.
  4. Identity and delegation. Ways to distinguish a person, the agent acting for that person and a swarm of subagents operating under bounded authority.
  5. Energy and compute accounting. Comparable reporting for the resource cost of training and long-running agent work.
  6. Human escalation. A common expectation that high-impact actions reach a responsible human with enough context to make a real decision.

This is not a treaty. It is plumbing for a world in which American and Chinese systems will compete, interoperate indirectly and appear together inside third-party products whether governments welcome the mixture or not.

The real Agent War is over defaults

The most important contest is not whose model wins one test in September 2026. It is who establishes the defaults for agentic work.

Who owns the identity layer? Who controls the tools? Where does memory live? Which model handles the expensive judgment step and which models perform the thousands of cheaper actions around it? Who can inspect the evidence later? Can a person or institution move its accumulated context to another system without beginning again?

MANGO-X begins with compute, enterprise access, frontier capability and integrated products. D-BAM begins with efficiency, open distribution, platform scale and intense internal competition. MiniMax's fifth chair reminds both sides that the interface is already expanding beyond text.

The meeting ends without a winner. That is the honest ending.

The next phase of AI will not be decided by one model, one country or one dramatic release. It will be decided by the ecology that makes intelligence useful, affordable, governable and persistent—and by whether humans retain enough visibility to choose among the machines gathered around the table.

The cast

  • MANGO-X — Microsoft: Copilot — Installed enterprise execution surface
  • MANGO-X — Anthropic: Claude — Deep research, writing and coding work
  • MANGO-X — NVIDIA: Compute foundry — Chips, systems and developer platform
  • MANGO-X — Google: Gemini — Research, cloud, data and distribution
  • MANGO-X — OpenAI: ChatGPT — General-purpose human workbench
  • MANGO-X — xAI: Grok — Live-world context and rapid infrastructure
  • D-BAM — DeepSeek: DeepSeek — Efficient reasoning and open deployment
  • D-BAM — ByteDance: Doubao — Consumer distribution and product velocity
  • D-BAM — Alibaba: Qwen — Open-model breadth and cloud reach
  • D-BAM — Moonshot AI: Kimi — Long-context research and agent work
  • Fifth chair — MiniMax: Multimodal systems — Text, speech, image and video creation

Source notes

Todd Kelsey · Tsunami Labs · September 2026