Contents
- Two kinds of AI agents are exposing automation’s limits
- Persistent threads and driverless robotaxi tests point to work control
- Technical constraints make AI agents impossible to automate without limits
- What data and channels should Vietnamese businesses test AI agents on
- Measure completed work before scaling AI agents
- References
AI agent are moving away from the chatbot format that answers one request at a time and into longer work chains. For Vietnamese marketers, this shift matters because performance no longer depends only on the model, but on data, connected tools and the action limits the business sets.
Key points
- Cursor Projects shows that an agent can maintain a workflow, coordinate subagents and proactively handle work over time.
- Driverless robotaxi operations show that real-world automation needs more layers of control than a model that can answer questions.
- Marketers must check access rights, input data and human approval points before scaling AI agents.
- The right metric is safely completed work, not the number of tasks an agent receives.
Two kinds of AI agents are exposing automation’s limits
In software, Cursor describes Projects as a way of working in which users do not open a new chat for every task. A coordinator agent maintains one thread, actively manages subagents and improves through use. This is a sign that agents are being designed as components that track work continuously, rather than just as question-and-answer windows. Source: Cursor.
In the physical world, Pony.ai and Verne robotaxi began fully driverless test runs on a 22 km public route to Zagreb Airport. NVIDIA said the vehicle uses NVIDIA DRIVE. These are two different environments, but they raise the same question: how far can a system act on its own while still controlling the output? Source: NVIDIA.
Persistent threads and driverless robotaxi tests point to work control
The update block below records only the capabilities described directly in the sources, then turns them into operational questions for marketing teams.
Cursor Projects: track one workflow instead of opening separate chats
Cursor Projects is a way to organize work around a long-lived thread. The coordinator agent can manage subagents and proactively handle tasks within the same project. For marketing teams, this touches brief management, content variants, review and handoff in one unified flow. Teams need to define clearly which documents the agent may read, which tools it may call and where human approval is still required. See Cursor’s description.
The 22 km robotaxi trial: testing automation on public roads
The Zagreb trial had passengers in the vehicle, no AV operator behind the wheel and ran on a 22 km public route to the airport. This is an operational test in an environment with external variables, unlike an agent processing data in software. Marketers can draw a similar requirement for automation: systems must be tested in real conditions, not only in demos. See NVIDIA’s information.
Technical constraints make AI agents impossible to automate without limits
Long-running work state: data must still be correct when the agent returns
Cursor Projects shows that an agent’s value is not only in answering a prompt, but in keeping the work moving and coordinating multiple subagents. Robotaxi, in turn, shows that an autonomous system must keep processing as conditions change on the road. Put together, the two sources make the first constraint clear: automation needs reliable state. If the brief, product price, inventory or access rights have changed but the agent still uses old data, continuous execution will spread errors across multiple steps.
That is why a marketing workflow needs a place to store document versions, rules for when data expires and logs showing which inputs the agent used. This is a more specific technical requirement than simply choosing a “smart” model.
Action rights: handling work does not mean deciding everything
The coordinator agent in Cursor can proactively manage subagents, while the robotaxi is described as operating without anyone sitting behind the wheel. Both show that the boundary between “doing the work” and “being allowed to decide” must be designed in advance. For marketing, an agent may draft copy, classify leads or prepare reports on its own; but sending mass email, changing budgets or editing customer data needs a stop point and a responsible human.
The broader the permissions, the greater the impact of an error. Businesses should split workflows by risk level, log actions and require approval before steps that affect the outside world.
What data and channels should Vietnamese businesses test AI agents on
These two developments are not proof that every business can deploy autonomous agents right away. They are evidence of a design direction: systems must be tied to a specific environment, data set and control mechanism.
In Vietnam, the main challenge often lies in data scattered across CRM systems, e-commerce platforms, ad platforms and internal files. Vietnamese data can also contain proper names, abbreviations and processing rules that differ across departments. For that reason, it is better to start with a narrow workflow such as generating lead reports or preparing ad variants. The agent should read only the necessary fields, produce output for the owner to check, and only then move on to API calls or system updates.
Teams also need to consider connectivity, access rights and vendor support. An agent may keep a thread well, but it still will not help if the CRM has no suitable API or the data does not have a stable format. The Zagreb experience also suggests that testing must happen under conditions close to real operations, rather than being judged only on a few sample prompts.
Measure completed work before scaling AI agents
- Choose one workflow with clear inputs, outputs and a responsible owner; do not start with a broad goal like “automate all marketing.”
- List the agent’s access rights: data it can read, tools it can call and actions that must go through approval.
- Test a real work chain with missing data, stale data and unusual inputs; keep logs to find errors.
- Only scale after measuring the share of usable output, time saved and the number of times humans had to fix it, instead of just counting how many tasks the agent received.
Cursor Projects and robotaxi are not in the same industry, but they make one principle clear for marketing: automation only has value when the system knows what it is doing, how far it is allowed to go and how to stop when conditions change. The model is only one part of the equation; data, integration and control decide whether a workflow can be used over the long term.
See more marketing analysis and guides at https://marketing365.vn.
Follow more analysis from Marketing365 to stay updated on the latest marketing trends.
Read more articles in the same category AI Developments.
