An agent (智能体) is halfway through booking a flight when the page reloads. It has the itinerary, the price, the passenger name, and no memory of which step it had reached. A human would sigh and start over. A model cannot even sigh.
That failure is not a reasoning failure. It is an environment failure, and it is the part of the agent stack that almost nobody benchmarks.
Key takeaways
- Tencent Cloud upgraded its Agent Runtime at its ecosystem conference on 29–30 October, centred on the sandbox that hosts each agent.
- Stress tests put sandbox start-up at 60 milliseconds and delivery at 600,000 sandboxes per minute for identical images, 160,000 for mixed ones.
- Snapshot, pause, resume and clone serve two masters: rollback during agent training, and recovery when a production task stalls.
- Tencent says the same runtime is used to post-train Hunyuan, MiniMax and StepFun models, which makes infrastructure a training input, not an afterthought.
- All figures are vendor stress tests; none has been independently reproduced.
What was actually announced
At the 2026 Tencent Global Digital Ecosystem Summit in Shenzhen, Tencent Cloud presented an upgraded Agent Runtime: the sandbox layer, the deployment tooling, state management, plus storage products it calls AgentCBS, AgentCFS and Agent Bucket, and an observability console.
The underlying sandbox (沙箱) technology, CubeSandbox, has been open-sourced and iterated through the year. The upgrade front-loads resource pre-creation, image warming and template snapshots, so that when a request arrives, the scheduler is not starting from zero.
The architecture is not exotic. What is notable is the framing. Ma Wenshuang, who leads Agent Infra at Tencent Cloud, described three different infrastructure demands arriving at once: high-frequency tasks need an environment immediately, long-running tasks need state preserved, and autonomous execution needs isolation and authorisation. Those are three different products in most stacks. Tencent is selling them as one.
Why a training loop cares about a sandbox
Here is the part that makes this a model story rather than a cloud story.
Reinforcement learning for tool-using agents requires the model to try things and see what happens. Terminal tasks need containers. Operating-system tasks need virtual machines. Browser tasks need a browser. Running thousands of these in parallel is expensive, and running them serially is hopeless — the model spends most of its wall-clock time waiting for an environment to come back.
Two capabilities change that arithmetic:
- Fast creation. If an environment takes a minute to provision, a rollout that fails in three seconds still costs a minute. At 60 milliseconds, failure is cheap enough to be part of the search.
- Snapshot and rollback. If the model can branch from a saved state, one exploration can spawn twenty variants without replaying the first half of the task.
Tencent says the runtime is already used in post-training for Hunyuan, MiniMax and StepFun models. That claim is worth pausing on. It means the sandbox is not a deployment convenience bolted on after training; it is in the training loop, and its latency shapes what kind of curriculum a lab can afford.
Reading the throughput numbers
Two figures circulate and they look contradictory: 600,000 sandboxes per minute, and 160,000 per minute. They are not a mistake.
The 600,000 figure applies to homogeneous images — every sandbox boots the same image, so caching and pre-warming do almost all the work. The 160,000 figure applies to heterogeneous images, where each request needs a different environment and the cache hit rate collapses. Tencent's own materials quote both, and the gap between them is a useful reminder that "start-up time" is a distribution, not a constant.
Tencent also claims that pause-and-resume, where an idle environment sleeps and wakes in milliseconds, cuts total enterprise deployment cost by 70%. That number depends entirely on duty cycle. An agent that is busy 90% of the time saves far less than one that waits on human approval for hours.
The security argument is the real one
Speed is the demo. Isolation is the sale.
The sandbox design puts each agent in its own kernel, restricts network access with ingress and egress rules, and — the detail that matters most — keeps sensitive credentials outside the sandbox, granting them per task. The model can run code and call tools without ever holding the company's long-lived keys.
For an enterprise, that is the difference between an agent that can be given a real workflow and an agent that has to be watched. Tencent's pitch leans on three things: every action logged, permissions scoped per task, and an environment that can be killed without killing the process that owns the credentials.
Tencent also reports that WorkBuddy, its AI workspace, now covers half of the Fortune China 500 and has processed 9 billion task requests. Those are adoption numbers, not reliability numbers, but they explain why the infrastructure argument is being made now.
Honest limitations
Every performance figure in this piece comes from Tencent's own stress tests, presented at its own conference. There is no third-party benchmark of Agent Runtime, and the 60-millisecond claim almost certainly refers to a warm, pre-created environment rather than a cold boot on an idle cluster.
The post-training claim — that Hunyuan, MiniMax and StepFun use the runtime — is Tencent's, and the three labs have not described how much of their training depends on it. I have not tested the sandbox, measured its isolation, or verified the 70% cost reduction, which is a modelled figure rather than a measured one.
Nor does any of this settle whether sandbox speed is the binding constraint on agent quality. A fast environment helps a model that already plans well; it does not fix a model that cannot plan.
What readers can do now
- If you run agents in production, measure your environment provision time separately from model latency. Most teams discover the sandbox, not the model, dominates tail latency.
- Ask your agent platform two questions: can it snapshot and resume a task mid-run, and where do credentials live during execution? The answers predict more about reliability than any leaderboard.
- Before adopting a vendor's sandbox numbers, ask whether they are homogeneous or mixed images. The two differ by nearly fourfold.