Big Tech AI Office Race: Why Harness Layer, Not Models, Determines Winners

Part 1: Current Capabilities — The Gap Is in the Harness Layer, Not the Model

Q1: How do leading AI office products compare? IDC says "complex tasks can be delivered but not directly used" — where is the bottleneck?

Adrop : The top tier consists of three players — Tencent WorkBuddy, ByteDance Doubao Work, Alibaba Qwen Office/DingTalk — plus Baidu Doudouzi/Wenku and Kingsoft WPS Lingxi. Their DNA differs: WorkBuddy evolved from the CodeBuddy coding agent, giving it a six-month head start; Qwen Office and Doubao Work only completed internal org restructuring and brand consolidation in July–August this year.

The reason products "can deliver but dare not be used directly" is not the base LLM but the surrounding Harness (driving engineering) . At similar parameter scales, model intelligence is roughly equivalent. The real differentiator is the outer engineering system: Is task decomposition accurate? Are tool-calling granularity and stability sufficient? Is long/short-term memory governance in place? Are there strict boundary constraints and result self-check mechanisms?

Often AI office "feels bad" because the Harness is undercooked: the agent runs for a long time, calls many tools, burns tokens, and delivers a half-finished artifact. Compared with overseas leaders like Cursor and Codex, domestic implementations still lag in low-level Agent Runtime engineering.

Q2: A test showed different vendors computing quarterly totals from the same Excel with million-yuan gaps. Is the model weak or is verification missing?

Adrop : I view such test conclusions with high skepticism. Excel statistics have no algorithmic black magic; automation scripts compute correctly, so the model isn't that weak. The huge variance stems from ambiguous prompts causing divergent decomposition. Different agents handle anomalous orders differently; once the call path diverges, numbers diverge. With precise prompt constraints, any mature agent can invoke a Python script to compute correctly.

This tests the agent's Harness driving ability: Can it decompose the task reasonably? Can it invoke a parsing script? Does it have built-in rules to verify the result? Can it self-heal when errors occur mid-run? Judging product quality by a non-transparent, fuzzy benchmark is unfounded.

Q3: Vendors are rapidly adding file browsers, canvases, multi-user collaboration. Features ship fast but feel unpolished — are fundamentals missing?

Adrop : Features will be copied to parity quickly; there is almost no moat. For example, WorkBuddy added a file browser supporting natural-language PPT generation with real-time right-panel tweaks — competitors can replicate it fast.

The perceived lack of solidity comes from the time needed to accumulate low-level engineering fundamentals. The underlying Agent Runtime — the execution environment that schedules interactions with the LLM — still carries legacy baggage in some teams; some haven't fully rewritten core components in high-performance languages like Rust. Overseas teams started earlier and iterate faster; domestic giants need time to mesh, but this is a pure engineering problem that time can close.

Q4: For generic tasks like writing docs, making PPTs, replying emails, models show no qualitative gap. Are vendors fighting the wrong battle?

Adrop : On generic tasks, differentiation is indeed minimal. The whole industry faces a fatal pain point — "AI flavor too strong" : mechanically symmetric short sentences, parallelism, two/three-part transitions, and forced grand endings that trigger visceral dislike.

Why are Coding Agents most successful? Because code has an objective, unique verification mechanism — does it compile, do tests pass? The machine decides. But writing quality is highly subjective and lacks automated verification. Piling flowery language on generic text is meaningless; the key is "remove AI flavor, add human flavor" — preserving the author's original voice and perspective.

Q5: WorkBuddy leads currently — is it product or ecosystem? How big is the pure product gap?

Adrop : Pure product feature gaps will vanish quickly; the essence is ecosystem war and entry-point war .

WorkBuddy does have concrete product-strategy advantages:

First-mover advantage & product experience : Ran half a year earlier, experience is more foolproof.

Multi-model access strategy : Cloud-side free switching across vendors (lightweight tasks route to ultra-cheap Zhipu GLM Flash or DeepSeek Flash, maximizing cost-performance), whereas Doubao Work currently binds mainly to its own model.

Connector & skill ecosystem : Backed by Tencent Cloud Skill Hub (China's largest skill distribution platform), it has accumulated massive connectors and expert skills, and pioneered an AI payment monetization loop via Alipay.

Part 2: Evaluating Each Vendor's Route — Without System Integration, AI Is Just a Chat Box

Q6: AI office competes on ecosystem. What does ecosystem mean here? Where is the deepest moat?

Adrop : Ecosystem here means the existing three collaboration bases: WeCom, DingTalk, Feishu (carrying OA, docs, collaboration, approvals, meetings).

Why fight on ecosystem? Because individual users don't pay; willingness to pay is near zero ; the real revenue is enterprise market. Enterprise migration cost for collaboration software is massive — they won't swap the entire base for a single AI office tool.

AI office is a "new entry point" growing out of the existing collaboration ecosystem , defending incumbency while capturing increment. Overseas competes on delivery completeness; domestic competes on ecosystem fit. The domestic moat runs deep in organizational relationship graphs — switching tools means rebuilding approval paths, permission systems, and knowledge deposits from scratch.

Q7: Everyone talks organizational context. Alibaba open-sourced a long-context tool — what's hard about this?

Adrop : Organizational context is high-value, the cornerstone of enterprise agents. The difficulty lies in data sources: historical enterprise documents and group-chat message history .

These two datasets are the deepest and most entangled inside enterprises. It tests not only knowledge-base retrieval but also the agent's comprehensive governance of permission boundaries, long/short-term memory, and employee role identities.

Q8: Byte plants AI inside Feishu, Alibaba open-sources context tools, Baidu pivots to vertical suites. How would you choose?

Adrop : Route depends on the cards each holds. Baidu lacks a collaboration base, so it must dig verticals like finance and legal. Feishu is a high-stickiness information-loop closure; Feishu data is Doubao Work's natural fuel. Qwen Office leans on DingTalk's huge B-end base and open-source ecosystem. The core is whether the product can grow on the real needs of the existing customer base.

Q9: Big companies can't know every industry, so they build open platforms. Why would developers put skills on your platform? What tangible value does the platform offer?

Adrop : The platform must deliver visible, tangible value. Taking WorkBuddy's open platform as example:

For individual/independent developers , it's a brand-new traffic entry and monetization channel. Previously a developer's scripts, prompts, or SOP workflows only lived locally or on GitHub, undiscoverable. Entering the Skill Hub and Expert Teams gives versioning and distribution entry; good work gets official exposure; the platform supports paid skills (per-call or subscription), closing the commercialization loop.

For enterprise developers , it's a connector to embed their own services onto the user's desktop. Enterprises can package mature industry software as Buddy Apps or connectors. For users, the max value is using natural language and agent context to connect heterogeneous systems (ERP, finance, CRM) for workflow orchestration. Hardware vendors also want to extend agents beyond the desktop. But the prerequisite: the platform must first generate strong ecosystem pull and provide transparent, stable rules.

Q10: How to vet external capabilities entering real tasks? Does the industry need a public exam standard?

Adrop : Leaderboard-style public exams mean little. Measuring external capability quality comes down to "how much friction when entering real tasks" :

Can it be precisely triggered by the agent in the right scenario when the user initiates a task?

Does it deliver a static, pretty half-finished artifact, or a real workflow where changing one number cascades through the whole process?

On failure, does it leave the developer clear, traceable logs, or does the user just silently uninstall?

Only capabilities that enter real business and close the loop on errors and feedback pass the real-world acceptance test.

Q11: WorkBuddy opened WeChat/WeCom connection capabilities to AI but didn't push open-source context tools. Given your knowledge of Tencent, how was this decided?

Adrop : Tencent internally has open-source Agent Memory components; technology isn't the bottleneck. The core is product strategy and extreme restraint on user privacy . In China, many people's work and life are deeply intertwined in WeChat and WeCom. Rashly connecting full group chats for memory would cross massive privacy/compliance red lines; DingTalk is purely work-oriented, making authorization toggles easier.

WorkBuddy focuses on connectors because an agent without external system connections is just a chatbot — it cannot do real work . Real work must produce deliverables: documents, code, reports. The agent must use MCP and connectors to link ERP, databases, email, moving and processing data from system A to system B — with connections, it becomes "doing work."

Part 3: Who Wins in Three Years — Don't Look at Three Years, Look at Fragmentation; It's About Who Outlasts Whom

Q12: If judging who wins in three years, which dimensions would you watch?

Adrop : Three years is too long! LLMs turn over every three months; you can't see three years out.

The domestic market will absolutely not produce a single "far ahead, winner-takes-all" outcome. The endgame will be fragmented, waxing and waning .

Everyone has an ecosystem base; features will level rapidly. The final contest is classic Tencent product philosophy: who makes no fatal product-strategy mistakes, who commits resources most steadfastly, who survives longest and outlasts opponents . In this marathon, fewer mistakes and longer survival beat short-term sprinting.

Q13: Everyone sells enterprise seats. Individual subscriptions or enterprise procurement — who pays? Is outcome-based pricing realistic?

Adrop : Individual paid rate is indeed extremely low (single digits or lower); individuals expect free, so C-end subscriptions can't support the market. The business center of gravity must be enterprise (seat fees + token fees).

But enterprise procurement is currently extremely cautious, because in the macro environment, enterprises can't calculate an ROI . SMB owners are pragmatic: buy AI seats, does revenue grow? Do customers increase? No one dares guarantee. AI can compress two days of grunt work to half a day, but the doers are often junior staff; enterprises won't directly lay off people, so the labor cost headcount doesn't materially drop. Cost reduction isn't the goal, and efficiency gains don't directly translate to revenue growth , so enterprise decision cycles are extremely long.

Currently, only three scenarios have genuine, strong willingness to pay with real money: coding (Coding Agents), AIGC visual short-form content production, and office . Office has the broadest user base but also the longest market-education and ROI-calculation cycle.

Q14: What's the next true game-changer technically? How to make deliverables directly usable?

Adrop : Only one metric matters: task completion rate .

The underlying framework remains: LLM + Harness . To make deliverables "directly usable," the key is five Harness loops:

Tool-call stability : Avoid mounting too many tools causing hallucinations and crashes.

Context governance : Cross-session long/short-term memory and unified role constraints.

Hard rule constraints : Strictly prevent the model from freewheeling.

Process & result verification : After generating output, independent rules or code verify correctness.

Correction & self-healing : When stuck or drifting mid-run, can the agent auto-retry and self-heal instead of blindly running to death?

Q15 & Q16: WeCom communication data is a trump card — why no big move yet? Why is turning data into corporate memory hard?

Adrop : The difficulty is not technical infeasibility; it's WeChat ecosystem's intertwined public/private nature and Tencent's habitual extreme restraint on touching user privacy data . WeCom connects colleagues, external clients, and WeChat friends, containing vast personal trivia and sensitive business dealings. Without absolute-safe industry-grade de-identification and user-authorization mechanisms, not touching this data lightly is prudent product restraint, not capability lack.

Q17: What important changes in big-tech AI office this half-year must be highlighted?

Adrop : Two points must be clear:

First, org integration is the watershed, but integration ≠ victory. Byte merged Feishu into Doubao; Tencent merged QClaw into WorkBuddy. Collaboration office is no longer an independent battlefield; it becomes "fuel" for the LLM entry point. But the merger is just stop-loss; post-merger tech-stack fusion and engineering landing still have a pile of pits to step in.

Second, enterprise selection advice: don't worship myths; choose by existing ecosystem. Heavy Feishu users → prioritize Doubao Work. Deep WeChat ecosystem, local-file intensive, with Xinchuang (IT innovation) requirements → prioritize WorkBuddy. Migration cost far exceeds the tiny product differences.

Whoever helps users actually push work into business systems is the one qualified to talk about ultimate moats.