Qwen and the Agent Frontier: What Multi-Step AI Means for Knowledge Work

For most of the past three years, the dominant question about generative AI has been a fairly narrow one: how good is the chatbot? Can it write a decent email, summarise a meeting, draft some code? That framing is now changing fast. The new question — the one Alibaba pushed firmly to the centre of the industry conversation at its annual Apsara Conference — is whether AI can plan and execute multi-step work on your behalf. Welcome to the agent frontier.

Alibaba's recent announcement of a comprehensive full-stack AI upgrade for the agentic era isn't just another model release. Paired with a new flagship Qwen model and the company's custom Zhenwu M890 chip, it's a bet that the next phase of AI competition will be won by whoever can stitch together silicon, models and orchestration into something that actually does things. For knowledge workers, that bet matters more than any single benchmark score.

From chatbots to agents: a meaningful shift

An AI "agent" is not a marketing rebrand of a chatbot. The difference is real. A chatbot waits for a prompt, produces an answer, and stops. An agent is given an objective — "prepare a competitor brief on these five companies and email it to the team by Friday" — and then plans, decomposes the task, calls tools (search, spreadsheets, CRMs, email), checks its own progress, and recovers from errors. It chains many model calls together rather than a single one.

This is the capability Alibaba is positioning Qwen around. As the South China Morning Post reported, the company is explicitly trying to become "China's AI factory" — and that factory metaphor is telling. Factories don't sell single artefacts; they produce work at scale. The same logic applies to agentic AI: the value isn't in one clever response, it's in the volume of routine cognitive work it can absorb.

Why silicon suddenly matters again

Running an agent is computationally very different from running a chatbot. A single user query might fan out into dozens of model calls — planning, retrieving documents, generating code, evaluating results, retrying. That's why Alibaba's hardware announcement is inseparable from its model story.

The newly unveiled Zhenwu M890, designed by Alibaba's T-Head semiconductor arm, is being pitched directly at NVIDIA's territory. According to Wccftech, Alibaba is claiming roughly three times the performance of NVIDIA's H20 with 144GB of HBM3 memory, and has laid out a roadmap extending through 2028. Digitimes describes the chip as a clear signal that Alibaba is doubling down on AI infrastructure rather than relying solely on imported silicon.

Whether those performance claims hold up in independent testing is a separate question. The strategic point is this: agentic AI is so memory- and bandwidth-hungry that any serious player needs vertically integrated infrastructure. If you're a knowledge worker wondering why Australian businesses keep hearing about chips when they ask about productivity software — that's why. The economics of an agent that runs all day, every day, are very different from the economics of a chatbot you poke at occasionally.

What this actually means for knowledge workers

Let's get concrete. If multi-step agents become reliable and cheap enough to deploy at scale, the work most exposed to change is the work made of connecting things: pulling data from one system into another, reconciling formats, chasing approvals, summarising long documents into shorter ones, drafting responses to predictable inputs.

For an Australian accountant, that might mean an agent that pulls client transactions, flags BAS anomalies, drafts the explanatory notes and queues a review. For a marketer at an SME in Brisbane, it might be an agent that monitors competitors, drafts campaign briefs, and schedules creative reviews. For a junior lawyer, it might be document review that previously took a week compressed into an afternoon — with the human stepping in to verify and decide rather than to assemble.

This is genuinely different from the "AI as autocomplete" experience of the last two years. It's also where the risks sharpen. An autocomplete that hallucinates wastes thirty seconds. An agent that hallucinates can send the wrong invoice to the wrong client, or commit your business to a contract clause you never reviewed.

Three questions worth asking before you deploy

For Australian businesses tempted to leap in, the agentic era requires a slightly different procurement mindset than the chatbot era. Three questions are worth pressing on:

A geopolitical undercurrent Australians shouldn't ignore

It would be naive to read Alibaba's announcement purely as a product story. The company is openly framing itself as the backbone of China's AI ambitions, and its custom-chip roadmap is at least partly a response to US export controls that have shaped what NVIDIA hardware Chinese firms can buy. The H20, which the Zhenwu M890 is being compared to, exists in its current form precisely because of those restrictions.

For Australian organisations, this matters in a practical sense: the agentic AI market is fragmenting along geopolitical lines. The model you choose isn't just a technical decision; it's increasingly a supply-chain and trust decision. Qwen models are open and capable, and many Australian developers will use them. But enterprise deployments will need to think harder than they did in 2023 about which stack they're locking into and what that implies five years out — particularly given Alibaba's chip roadmap stretches to 2028.

The honest middle path

The temptation in any AI cycle is to oscillate between two unhelpful poles: the breathless take that knowledge work is over, and the dismissive take that nothing has really changed since GPT-3.5. Neither is right. What's actually happening is that the unit of useful AI output is shifting from "a response" to "a completed task," and the infrastructure to support that — models, chips, orchestration layers — is being built in public, very quickly, by a handful of well-resourced players.

For knowledge workers, the right posture is curiosity with guardrails. Try the agents. Give them small, contained tasks. Pay attention to where they fail, because the failure modes tell you more about the technology than the demos do. And don't confuse the speed of vendor announcements with the speed at which any of this becomes safe to bet your business on.

The agent frontier is real. It's also a frontier — which means it's largely unmapped, and the first travellers will spend as much time learning what doesn't work as discovering what does.

Related on Bleen

Sources