Every now and then, a survey paper appears that does more than summarize a research field—it attempts to define where that field is heading.
The recently released survey Toward Edge General Intelligence with Agentic AI and Agentification: Concepts, Technologies, and Future Directions is one of those papers. Spanning more than one thirty pages and hundreds of references, it offers a comprehensive overview of one of the fastest-evolving areas in computing: the convergence of Edge AI, agentic AI, and autonomous systems.
Whether you agree with all of its terminology or not, the paper makes one thing very clear:
The conversation is no longer about deploying AI models at the edge. It is about building intelligent systems that can operate autonomously at the edge.
From inference to autonomy
For years, Edge AI has largely focused on a well-defined problem: how to execute machine learning models efficiently outside the cloud.
This led to tremendous progress in:
- model compression
- quantization
- hardware accelerators
- distributed inference
- resource-aware scheduling
These remain important research topics.
However, the survey argues that they are no longer sufficient.
Future edge systems will increasingly need to perceive their environment, reason over context, plan actions, collaborate with other devices, and continuously adapt to changing conditions. In other words, they must exhibit characteristics that go beyond inference and move toward autonomous decision making.
The rise of agentification
One concept that appears throughout the paper is agentification.
Rather than viewing AI as a model that receives an input and produces an output, agentification treats AI as an active software component capable of interacting with its environment, maintaining goals, invoking tools, communicating with other agents, and adapting its behaviour over time.
This distinction may sound subtle, but it fundamentally changes how we think about system design.
Instead of asking:
“Which model should I deploy?”
we increasingly ask:
“How should multiple intelligent components cooperate to achieve a long-term objective?”
That is a systems question as much as it is an AI question.
Edge General Intelligence
Perhaps the most ambitious concept introduced in the survey is Edge General Intelligence (EGI).
The idea is not to replicate Artificial General Intelligence on tiny devices. Rather, it envisions distributed edge infrastructures where intelligence emerges from the interaction of multiple specialized agents operating across heterogeneous devices and networks.
Whether “Edge General Intelligence” becomes the accepted terminology remains to be seen.
Nevertheless, the underlying direction is difficult to dispute.
Edge devices are becoming increasingly capable.
Small language models are becoming increasingly efficient.
Communication infrastructures continue to improve.
Taken together, these developments enable forms of distributed intelligence that would have been unrealistic only a few years ago.
A systems perspective
One aspect I particularly appreciated is that the survey does not focus exclusively on AI models.
It also highlights many of the systems challenges that will ultimately determine whether these visions become reality:
- orchestration across heterogeneous devices
- resource management
- communication efficiency
- trust and security
- privacy preservation
- collaborative decision making
- lifecycle management
- scalability
This is encouraging because it reinforces an important message that sometimes gets overlooked:
better models alone will not create autonomous edge systems.
Progress will depend equally on advances in distributed systems, networking, operating systems, hardware architectures, and middleware.
Final thoughts
Survey papers often serve as snapshots of a research field.
This one feels different.
It captures a transition that many of us have been observing over the past couple of years: the shift from Edge AI as efficient inference toward Edge AI as autonomous, adaptive, and collaborative intelligence.
Whether we eventually call it Edge General Intelligence, Agentic Edge AI, or something else entirely is perhaps less important.
What matters is that the research questions are changing.
And for those of us working at the intersection of AI systems, networking, and embedded intelligence, that makes the next few years particularly exciting.
Source: https://arxiv.org/pdf/2508.18725
