The Global Agent Race: Why Everyone Is Building Their Own OpenClaw
What This Is About
In the first article of this series, I argued that AI agents are becoming a new computing layer. In the second, I focused on the most concrete version of that shift: personal agents that already run on real hardware and do real work.
This third piece zooms out.
The interesting question is no longer whether agents work. They do. The more important question now is why ByteDance, Tencent, Baidu, NVIDIA, Abacus AI, and others all moved toward the same architecture at almost the same time.
My answer is simple: OpenClaw did not just prove that agents are possible. It made visible a new control layer between the model and the work.
That is the layer everyone now wants to own.
There is a familiar pattern in tech: an open system shows what is possible, then larger players rush in to package it, distribute it, secure it, or absorb it into their own stack.
That is exactly what is happening here.
OpenClaw started as an open-source agent framework. But what it really revealed was a more important idea: the next interface may not be a chat window at all. It may be a persistent software actor with memory, tools, permissions, and a channel into the systems where work already happens.
Once that becomes obvious, the strategic logic changes fast.
You are no longer building „a chatbot with more buttons.“ You are building the operational layer that sits between a model and email, calendars, files, documents, messaging apps, cloud services, internal tools, and eventually entire workflows. Whoever owns that layer owns a very privileged position.
That is why this is starting to look less like product experimentation and more like a land grab.
This Is Not One Race. It Is Several Different Bets
The important thing is not that everyone copied the same features. They did not. What we are actually seeing is several distinct strategies converging on the same architecture.
1. The operator bet: Abacus AI removes the setup pain
Abacus AI is making the most pragmatic bet of the group.
Its product, Abacus Claw, is a hosted, managed version of OpenClaw. The documentation is explicit about the value proposition: no local setup, no Node.js wrangling, no server maintenance, no babysitting of keys and infrastructure. According to Abacus, setup takes under a minute. The agent runs 24/7, carries persistent memory across sessions, includes a cloud Linux computer with terminal and browser access, and can operate across WhatsApp, Telegram, and Slack at the same time.
That may sound boring next to the louder agent demos. It is not boring. It is strategically important.
OpenClaw’s biggest current weakness is not capability. It is friction. Self-hosting is still a power-user activity. Abacus is betting that the first company to remove that friction without removing the agent’s usefulness gets a meaningful share of the market.
That is a serious bet, because this is how technologies usually spread. Not where they are invented, but where they become easy enough to survive first contact with normal humans.
2. The distribution bet: Tencent wants the agent to live inside the message stream
Tencent’s move is different and, in some ways, more ambitious.
Its official QClaw product page describes an AI agent from Tencent PC Manager that users can deploy locally on Mac or Windows, bind directly to WeChat, and use to control files, browser tasks, email, GitHub workflows, reminders, and research jobs through ordinary messages. The positioning is clever: not „install a framework,“ but „send a WeChat message and let the agent handle it.“
That matters because Tencent is not just shipping a tool. It is trying to place the agent inside the communication layer where work already happens.
A Reuters report from March 22 described Tencent’s WeChat integration as a contact-level agent interface inside China’s dominant messaging platform, a product with over a billion monthly active users. Tencent Cloud is pushing the same logic on the infrastructure side, offering one-click OpenClaw deployment and highlighting support for WhatsApp, Telegram, Discord, Slack, WeCom, QQ, DingTalk, and Lark.
This is the deeper pattern: Tencent is not treating agents as a side feature. It is treating them as a native extension of messaging, desktop control, and cloud distribution all at once.
If your agent lives inside the place where messages, files, approvals, and coordination already flow, that is not an add-on. That is a bid to make the messaging layer itself the operating system for work.
3. The ecosystem bet: ByteDance is wrapping the framework in its own stack
ByteDance is pushing from another angle.
Volcengine’s ArkClaw documentation now includes one-click access for ArkClaw, updated in late March. A 36Kr walkthrough reported that a cloud-based ArkClaw instance could be created in roughly two minutes, then connected to ByteDance products like Feishu and Anygen. In other words, ByteDance is not only hosting the framework. It is wiring it into its own office software, agent layer, and model stack.
That is a stronger move than simple hosting.
Caixin reported in early April that Volcengine had officially sponsored OpenClaw and planned to co-build the Chinese mirror of ClawHub, the skill repository around the ecosystem. The same report said ArkClaw was being integrated with Doubao models and Feishu, while Volcengine had reached 140 enterprise clients and crossed the 1 trillion token threshold by March 2026.
That is not hobbyist energy anymore. That is platform energy.
ByteDance seems to understand that if agents become real infrastructure, the valuable position is not just „we also have one.“ The valuable position is to be the place where agent runtime, collaboration software, skills, models, and enterprise distribution already fit together.
4. The trust-layer bet: NVIDIA is trying to own the safety architecture
NVIDIA is making a very different, and in my view underrated, bet.
Its OpenShell and NemoClaw announcement is not mainly about distribution. It is about trust.
NVIDIA’s argument is blunt: long-running agents are useful, but current runtimes are missing the security primitives needed to trust them. OpenShell tries to solve that with out-of-process policy enforcement, isolated sandboxes, deny-by-default controls, audit trails, and a privacy router that can keep sensitive context on-device while selectively routing to outside models.
That is the right problem to attack.
The current wave of agent enthusiasm routinely skips over the ugly bit: once an agent has persistent shell access, live credentials, tool-install rights, memory, and six hours of context, it stops looking like a chatbot and starts looking like a new attack surface with excellent manners.
NVIDIA sees that. More importantly, NVIDIA is building for the buyers who see it too.
If OpenClaw proved that agents can work, NVIDIA is trying to provide the answer to the next enterprise question: „Fine. But how do we let them work without giving our security team a collective nervous twitch?“
That is not a small niche. It may become one of the decisive layers in the whole market.
5. The convergence signal: even the search and platform incumbents are joining in
Baidu matters here less because its exact product design is the most interesting, and more because its presence confirms the broader convergence.
Reuters reported in March that Baidu introduced new AI agents designed for multi-step tasks such as editing videos, creating presentations, conducting research, and even ordering coffee. Bloomberg also pointed to Baidu integrating OpenClaw-style agent capabilities into its Xiaodu device ecosystem.
The important part is not the coffee. It is the pattern.
When search companies, messaging giants, cloud vendors, model providers, and infrastructure companies all move toward persistent tool-using agents at the same time, you are no longer looking at a quirky open-source subculture. You are looking at the early formation of a market.
What All of These Companies Actually Understood
Here is the simplest way I can put it.
A model answers. An agent observes, remembers, decides, uses tools, and keeps going. An agent platform controls the environment in which all of that happens.
That last layer is where the real leverage sits.
Who stores the long-term context? Who controls permissions? Who owns the integrations? Who defines the safety boundary? Who is the default channel the user reaches for when they want work to get done?
Those are not feature questions. They are control questions.
That is why this matters more than another round of AI branding.
The contrarian view
A little cold water is healthy here.
First, a lot of this is still packaging, not proof. Product pages are not the same thing as reliable production use. The real test is what happens after week three, when the agent has to survive permissions, edge cases, broken integrations, confusing user requests, and the kind of chaos that real work produces on a Tuesday afternoon.
Second, many of these systems are converging on the same demos because the demos are easy to understand: schedule a task, summarize a document, answer a message, automate a workflow. The harder questions are still mostly unresolved: governance, observability, auditability, prompt injection, token economics, and operational failure modes.
Third, open source still has more strategic power here than the bigger players would probably like. OpenClaw did not just create demand. It also made the category legible in public. That makes it harder for any one vendor to fully lock down the shape of the market.
So yes, this is a race. But it is still an open race.
What I would actually watch from here
If I had to evaluate the field today, I would watch five things.
- Distribution: Does the agent live in the messenger, the cloud console, the desktop, or the enterprise workflow stack?
- Friction: How quickly can someone go from curiosity to a working agent that survives real use?
- Trust layer: Are permissions, isolation, logging, and policy enforcement real, or just decorative?
- Context ownership: Who keeps the memory, and how portable is it if the user wants to leave?
- Economics: Do the numbers still work when agents run continuously instead of as short-lived demos?
The companies that answer those five questions well will not just have a nice AI feature. They will control a meaningful slice of the next software layer.
The bigger picture of the series
Part 1 of this series argued that agents are becoming the new computing layer.
Part 2 showed that this layer already works on personal infrastructure, with all the promise and all the messiness that implies.
Part 3 adds the next obvious conclusion: once a new layer works, the race to control it begins immediately.
That is what we are seeing now.
Not a random pile of agent launches, but multiple serious bets on who gets to own the layer between models and work.
In the next article, I will go one step further into the uncanny part: self-improving agents, systems that do not just execute tasks but iteratively improve the way they execute them.
That is where the story stops feeling like software evolution and starts feeling like a species upgrade with paperwork.
Sources
Previous articles in this series
- The New Computer: Navigating the Global AI Agent Arms Race
- Personal Agents: When Your AI Actually Shows Up for Work
Hosted and operator layer
Tencent and messaging-layer distribution
- QClaw official page
- Tencent Cloud: Deploy OpenClaw in Seconds
- Reuters: Tencent integrates WeChat with OpenClaw AI agent amid China tech battle
ByteDance and ecosystem integration
- Volcengine ArkClaw documentation
- 36Kr: ByteDance’s ArkClaw May Be More Suitable for Ordinary People
- Caixin Global: ByteDance’s Volcengine Powers AI Growth with OpenClaw Partnership
Kommentar abschicken