Generative AI, including popular formats like ChatGPT, Copilot, and other chatbots, is based on large language models (LLMs). Agentic AI uses LLMs as a starting point to autonomously access and act on internal and external resources such as databases, financial accounts and transactions, travel services, etc.
Wikipedia defines AI agents
as:
“… a class of intelligent agents distinguished by their
ability to operate autonomously in complex environments. Agentic AI tools
prioritize decision-making over content creation and do not require human
prompts or continuous oversight.”
Aaron Holmes, in an article
for The Information, described seven kinds of AI agents:
“…business-task agents, for acting within enterprise
software; conversational agents, which act as chatbots for customer support;
research agents, for querying and analyzing information (such as OpenAI Deep
Research); analytics agents, for analyzing data to create reports; software
developer or coding agents (such as Cursor); domain-specific agents, which
include specific subject matter knowledge; and web browser agents (such as
OpenAI Operator).”
Agentic AI is given the
capacity to act independently of user supervision. Several protocols have been
developed to facilitate agent/user interactions and agent-to-agent interactions
Amazon scholar Michael Kearns, who is
also a professor of computer and information science at the University of
Pennsylvania, answers three questions for The Guardian:
1. What will be some of the challenges with entrusting
agentic AI with consequential actions like accessing sensitive data, or
autonomously making financial transactions?
One
challenge he notes is that LLMs don’t work with words, sentences, and images
but rather with embeddings, which are abstract representations of
words, sentences, and images.
“Embeddings preserve the meaning of content at the
expense of losing some details, like the exact sequence of every word in a long
novel. Thus, when content or context are shared across agentic AI systems,
drawing precise boundaries around sensitive or private information like
financial data will require careful handling.”
2. Given the fact that so much of the activity in large
enterprises and projects happens informally between humans, how will we
transition such interactions to an agentic world?
He answers that specifying
and enforcing what might be called “agentic common sense” will be
required. He suggests that it will take time and training for agentic AI to be
able to distinguish and emulate common human informal behaviors, those
common-sense behaviors that humans do.
“…users and enterprises should again be explicit in the
constraints they want agentic AI to obey and ask the AI to check with humans in
the case of any ambiguity.”
3. Sometimes we will want agentic AI to not just execute
transactions on our behalf, but to negotiate the best possible terms. Where can
we look for guidance as to how this might play out?
He notes that we have already
done this with behavioral economics and game theory.
“While generative AI seems to have arrived
instantaneously and matured rapidly in just the last few years, it in fact is
firmly grounded in decades of foundational science in areas such as machine
learning and neural networks. With agentic AI, we are in genuinely new
territory, with few established scientific and engineering principles to set
expectations. The challenges discussed here – privacy boundaries in embedding
space, agentic common sense and agent-to-agent negotiation – are a few of the
scientific hurdles that must be sorted out as AI becomes more autonomous. In
the meanwhile, leaders should encourage their organizations to be explicit
about the constraints they want agentic AI to obey and try to articulate their
enterprise common-sense policies.”
Turning AI into autonomous
agents comes with potential advantages as well as potential pitfalls that must
be overcome.
Building Agentic AI Trust and Transparency
Clarke Rodgers of Amazon Web
Services notes that agentic AI could potentially “mishandle sensitive data
or make high-stakes decisions without human input,” which could be
problematic. He suggests there is a need to redefine the intersection
of autonomy, transparency, and security. This involves putting strict
boundaries on agentic AI’s decision-making authority.
“A common approach is embedding
"human-in-the-loop" or "human-on-the-loop" frameworks,
which make sure that autonomous agents are either supervised or escalated to
human reviewers for critical decisions.”
An example he gives is
allowing agentic AI to flag suspicious account activity, but it requires humans to
review and make decisions based on that flagging. He notes that there is a need
for clear explanations of AI agents' capabilities and limitations. This aids
what he calls transparency as a strategy.
Regarding security, he
mentions the importance of secure-by-design architectures.
“This includes limiting access to sensitive systems,
real-time monitoring of AI behavior and applying zero-trust principles that
treat AI agents as potentially risky actors until proven safe.”
This often involves probing
agents for weaknesses just as an adversary might do. Thus, these simulations
help to prepare agentic AI for such issues.
Aside from risk mitigation,
there is a need to build trust in agentic AI systems among users of those
systems. One way to do that is by:
“…implementing secure feedback loops where user input
helps fine-tune agent behavior through real-time satisfaction ratings,
sentiment analysis and compliance-filtered interactions, organizations are
enhancing customer experience while maintaining strict security controls and
regulatory compliance.”
Agentic AI Expands AI’s Capabilities: Kearns Explains
AI agents are also designed
to learn from all their interactions with humans and with other agents. The
agent often acts as a capable human assistant, which frees up time for the
human(s) it works for. These capabilities are not only in language but also in
coding, mathematical reasoning, optimization, workflow planning, and many other
tasks. He notes that modern technology is often human-centric, but agentic AI
utilizes what are called ‘native languages’, such as embedding or embedding
space.
“Embeddings are an abstraction that provides great power
and generality, in the form of the ability to represent not the literal
original content (like a long sequence of words) but something closer to its
underlying meaning. The price for this abstraction is loss of detail and
information.”
That loss of detail needs to
be mitigated through further training. Agents interact with humans, content,
and other agents, all of which help them continue to train and refine.
One problem he brings up is
that embedding ‘languages’ differs among companies and is often considered
proprietary. It would be better, he says, if there were standardization. This
would not need to be total, and some proprietary information could still be
retained. He advocates for a common ‘base embedding’ that would be beneficial.
Context can be thought of as
the “working memory” of the LLM, and like humans’ working memory, it can be
selective and imperfect. That working memory becomes crucial for agentic AI
applications.
“How will context and its limitations affect agentic AI?
If embeddings are the language of LLMs, and context is the expression of an
LLM’s working memory in that language, a crucial design decision in agent-agent
interactions will be how much context to share. Sharing too little will
handicap the functionality and efficiency of agentic dialogues; sharing too
much will result in unnecessary complexity and potential privacy concerns (just
as in human-to-human interactions).”
Kearns explores agentic AI’s
ability to find bargains and to bargain with other ideas like game theory and
behavioral economics. Behavioral economics has shown that humans often act
irrationally when making purchases and decisions. Thus, what one should do
often does not sync with what one does. He gives an example of the ultimatum
game in game theory. He suggests that LLMs already mimic such odd human
economic behavior to some extent, and that will become more important as
agentic bargaining becomes more common and we delegate more bargaining power to
AI agents.
Next, he comes back to the
idea of agentic common sense, noting that AI has struggled with developing this
common sense since its beginnings. The sheer information size of the internet
and AI’s access to it has made developing a generic AI common sense successful.
However, more needs to be done for it to develop common sense that is specific
to us, what he calls subjective common sense. These are issues around trust and
security.
He concludes that agentic AI
is still very new to most users but is growing quickly, just like generative AI
did.
“The agentic-AI era is in its infancy, but we should not take that to mean we have a long and slow development and adoption period before us. We need only look at the trajectory of the underlying generative AI technology — from being almost entirely unknown outside of research circles as recently as early 2022 to now being arguably the single most important scientific innovation of the century so far. And indeed, there is already widespread use of what we might consider early agentic systems, such as the latest coding agents.”
References:
The
science of agentic AI: What leaders should know. Michael Kearns. Amazon scholar.
The Guardian. October 27, 2025. The
science of agentic AI: What leaders should know | Business briefs | The
Guardian
Scientific
frontiers of agentic AI. Michael Kearns. Amazon Science. September 11, 2025. Scientific
frontiers of agentic AI - Amazon Science
The age of agentic AI: Building trust and transparency. Clarke Rodgers. The Guardian. September 19, 2025. The age of agentic AI | Business briefs | The Guardian
AI
agent. Wikipedia. AI agent -
Wikipedia
The
Seven Kinds of AI Agents. Aaron Holmes. The Information. July 7, 2025. The
Seven Kinds of AI Agents — The Information
No comments:
Post a Comment