By Lex Chen, CTO Advisor at Atlas
On August 3rd, Alibaba dropped two big announcements at once: the new flagship model Qwen3.8 (2.4 trillion total parameters) went live, with major gains in coding and professional office (Cowork) capability; and at the same time, QwenWork — an enterprise-grade agent product formed by merging QoderWork, MuleRun, and Wukong — opened its public beta, billed as the industry’s first product to cover desktop agents, cloud agents, and enterprise-collaboration agents all at once.
The buzz is real — leaderboard rankings, a stock pop, open-source plans, deep DingTalk integration. But as someone who advises Atlas on where to place its own bets in this space, the question I actually care about is: does this launch change anything for how we should be thinking about AI at Atlas?
An unglamorous conclusion, up front
Here’s the conclusion first: this category of AI office agent has already reached a high degree of feature parity. Whether it’s Alibaba, Tencent, ByteDance, or Anthropic and OpenAI, everyone is essentially doing the same thing — wiring a large model into your docs, spreadsheets, IM, and email so it can write, look things up, and organise for you.
What actually determines whether a tool works well was never “which vendor added which button.” It comes down to four things:
- the user’s own level of judgement;
- the model’s underlying capability;
- how deeply the tool is integrated into the office system;
- how complete the company’s internal knowledge base is.
That fourth point is the one people underrate — and it’s the one most relevant to us. If every individual has to rely on their own tool to “understand” the company’s documents from scratch, you get massive, redundant token spend, and no guarantee of consistent quality. It’s not that the model isn’t smart enough; it’s that the same institutional knowledge gets relearned, slightly differently, by every single person who touches it.
So for a company like Atlas, the thing worth investing in was never chasing whichever specific product is trending. It’s two more fundamental things: building an agent-friendly office environment and enterprise knowledge base, and cultivating “super individuals” who actually know how to drive an agent well. Get those two right, and the company can swap in better models and newer tools quickly, and actually fuse agent capability with the real business — instead of just following the news cycle.
What is a company actually trying to achieve by rolling out AI to employees?
Set the specific product aside for a moment. When a company equips employees with this kind of tool, there are really only three effects it could be going for — and the cost and payoff of each path are completely different.
Option 1: give everyone a seat. This is basically assigning every employee a personal assistant. What we mostly see in practice: deliverables come faster, look richer, feel more “professional” — but the actual outcomes don’t get better. Sometimes they get worse, because the output now exceeds the user’s own ability to judge it. The ceiling on what an assistant can do is set by the person directing it.
Option 2: employees drive it themselves. The company pays for results and rewards the stronger individuals who take on more responsibility with pay, performance bonuses, and token subsidies. Underneath, this is really the company hunting for a new kind of talent — where “job capability” is now the combination of personal ability plus AI leverage.
Option 3: build intelligent support into enterprise operations. Embed agents into the supporting work of functions like HR, finance, audit, compliance, and admin, directly inside the digital workflow. This also opens the door to a more objective, third-party lens for evaluating collaboration quality and work output — a fairer, more thorough, more professional form of performance review.
Right now, this launch included, most of the big players — Alibaba, Tencent, ByteDance, OpenAI, Anthropic — are mainly betting on Option 1. But plenty of companies that have actually tried it find the ROI disappointing, and that tracks with an obvious pattern: most employees aren’t short on output — they’re short on the ability to do things correctly, and the ability to scale their own successes across a team. Handing a stronger digital assistant to someone who can’t drive it doesn’t produce better results; it just produces more output that no one can steer.
Option 2 is genuinely happening in the real world. But since most managers still don’t have an objective performance framework, the value of AI usage ends up wildly uneven across functions — it shows up clearly in a few key roles, while for most people there’s no scenario that makes the value visible, which makes it hard to reward the right behaviour consistently or push adoption company-wide.
Option 3 — a systematic agent platform that gives everyone digital capability while also managing and auditing the work process — is the hardest to pull off. Think: engineers required to submit code that an agent audits and enforces standards on; customer service reps whose day-to-day feedback gets automatically triaged by an agent, with tone and service quality reviewed. This requires tight integration with the actual business, non-trivial adaptation to how the org is structured, and dedicated staff to drive it. Every major vendor touches this today, but all of them are still scratching the surface.
Where QwenWork is actually pointing
On the surface — web app, desktop client, free/personal/enterprise tiers — QwenWork still looks like Option 1: give everyone a smarter assistant. But look closer at a few of its moves and it’s clearly leaning toward Option 3:
- It’s a merger of three separate products (QoderWork, MuleRun, Wukong) into one, spanning desktop, cloud, and enterprise-collaboration agents at once. That’s not the move you make if all you want is “a chat assistant for individuals.”
- It has deep DingTalk IM integration, and explicitly says it plans to connect to a company’s actual databases and workflows going forward — that’s the exact pivot from “personal productivity tool” to “enterprise operations backbone.”
- Its pricing splits into separate personal and enterprise tracks, hinting at two different underlying products: the personal tier sells efficiency, the enterprise tier likely sells process control and visibility into the work itself.
This confirms the earlier point: selling “a smarter assistant” on its own has a fairly low ceiling. The real upside is whether an agent can actually be woven into a company’s real workflows and knowledge systems, becoming part of how collaboration gets objectively evaluated. Whoever nails that first is the one who actually cashes in the value of agents — instead of just making deliverables look more polished.
Back to Atlas: should we lock in on one tool?
My take is still no.
This category of tool isn’t ready to be locked in on for the long haul, and there’s no way — even at the company level — to reliably pick a single globally optimal product. The more realistic move is to encourage our team to keep experimenting and settle on whatever tool they personally use well, then put the real effort into sharing good scenarios and usage experience — not standardising around one specific skill or one product’s particular feature set. Three reasons:
- These products iterate too fast, and they’re all borrowing each other’s best features. The actual gap between them isn’t that large, but each one is evolving quickly — QwenWork itself is proof: six months ago, it was three separate products.
- The “best model” title keeps changing hands, with every vendor trading places at the top. Locking into one tool can’t guarantee long-term superiority, and that one variable happens to matter enormously for how well the tool actually performs.
- The biggest cost in this whole toolchain is still token spend — and that’s not just the per-million-token sticker price, it’s also model capability and whether you’re using it correctly. Getting to true cost-optimal usage can’t be solved short-term by “one standard for everyone.” It only comes from individuals putting in the work to figure out the trade-off between cost and quality themselves.
Where this does matter directly for Atlas is that fourth point from earlier: the knowledge base. We’re already building in this direction — Florence, our internal knowledge base AI agent, draws on Atlas’s own API documentation, logs, processes and data rather than making every team member relearn it from scratch. That’s the underlying principle QwenWork is chasing at enterprise scale, and it’s the one piece of this whole conversation that’s genuinely worth investing regardless of which vendor is trending this quarter.
Qwen3.8 and QwenWork are, on the whole, good news for the industry — performance moved forward again, and prices dropped further. But for us, what actually matters was never who announced what this week. It’s whether we’re continuously building our own ability to drive an agent well, and whether Atlas is seriously building the knowledge infrastructure that lets an agent genuinely understand our business. Tools will keep changing. Those two things won’t.
