A fleet of AI agents large enough to catch the eye of independent researchers is running on Tencent’s infrastructure. That same fleet is repeatedly querying Amap, the mapping service owned by Tencent’s biggest rival, Alibaba. Hold those two facts next to each other and you get the shape of the story: something is operating at scale across competing Chinese tech stacks, and nobody outside the operation seems to know what it’s for.
Independent researchers posted preliminary findings on Sunday. BizToc picked it up Monday, October 5, 2026, and the coverage has been spreading since. The research is still ongoing, which means most of what follows is observation rather than conclusion. But as someone who spends his days sorting working agents from demo-ware, I find the observations more interesting than any tidy explanation would be.
How you find a fleet nobody announced
The detection method is the part I keep thinking about. The researchers weren’t watching Amap. They were watching traffic to URLquery, a domain-scanning service that AI agents commonly use to load websites they can’t reach directly. That’s the agent equivalent of a service entrance, and it turns out you can learn a lot by standing next to one with a clipboard.
If you build agents, you already know why this works. Agents hit walls constantly: geo-blocks, bot detection, JavaScript-heavy pages, rate limits. A scanning or proxy-style service becomes the workaround of choice, and that workaround leaves a trail. The agents weren’t hiding especially well, because nothing in a typical agent pipeline is designed to hide. It was designed to get the page.
The practical takeaway for anyone shipping agents right now: your traffic patterns are a signature. If your agent routes through a shared intermediary, your activity is visible alongside everyone else’s activity through that same intermediary. You are not anonymous. You are in a crowd, and crowds are exactly what researchers monitor.
Parallel, not coordinated
Here’s the detail that separates this from the swarm architectures everyone has been building toward. The agents appear to operate in parallel with no communication between them. No shared memory, no message bus, no orchestrator handing out tasks and collecting results.
That’s a design choice with real tradeoffs, and it’s worth understanding why someone would make it:
- It’s cheap and simple. Spinning up many identical agents with the same instruction requires no coordination layer, no consensus logic, no state synchronization. You just run the thing many times.
- It’s fault-tolerant by accident. If one agent fails, nothing else notices. There’s no single point of failure because there’s no single point of anything.
- It’s wasteful. Without shared state, agents duplicate work. The same query gets asked repeatedly because no agent knows another already asked it.
- It’s hard to attribute. Independent agents with no central controller look less like one operation and more like background noise.
Most of the agent frameworks I review push hard in the opposite direction, toward orchestration, shared context, and careful task delegation. This fleet suggests that for some workloads, brute parallelism with zero coordination is good enough. That’s a useful reality check for teams building elaborate multi-agent plumbing before they’ve confirmed they need it.
The queries themselves
What are they actually asking? Directions to entrances of public places. Parks, zoos, hospitals. Not addresses, not routes between cities, not business listings. Entrances.
I’m not going to speculate about intent, because the researchers haven’t concluded anything and I’d be making it up. But I’ll note what kind of task this resembles. Entrance-level detail is the sort of thing that’s thin or missing in most map databases, and filling that gap is a classic data-collection job. Agents are good at exactly this: a narrow, repetitive query pattern run thousands of times against a source that already has an API-shaped answer.
Whether this is dataset building, service testing, or something else entirely is unresolved. The honest answer is that nobody reporting on it knows yet.
What this tells us about the current state of agents
The researchers’ broader point lands harder than the specifics: persistent AI agent activity is now a permanent feature of the internet. Not a pilot, not a demo, not a scheduled batch job. Fleets running continuously, generating enough traffic that outside observers notice them through a third-party scanning service.
For anyone operating a web property, that changes the baseline. A meaningful slice of your traffic is agents doing tasks on behalf of somebody, and you probably can’t tell which agents or whose tasks. For anyone building agents, it means you’re operating in shared space where your behavior is measurable by strangers.
The fleet asking for zoo entrances may turn out to be mundane. The fact that researchers found it by watching a side door is the part worth filing away.
đź•’ Published: