AGENTIC

The Rise of AI Agents

About this document
This paper provides an overview of the growing role of AI agents for
established work functions, through definitions, types of agent, examples of
existing agentic applications, future application of AI agents and their
challenges and risks. A NotebookLM audio conversation version of this paper
is available at this link. This paper was written by James Carson.

Key section headings:

What are AI agents, and what is the ‘Agentic’?​ 3
5 types of AI agent​ 6
Existing Agentic tools​ 7
Applications of AI agents​ 11
Challenges and risks​ 14
4 viewpoints from Big Tech​ 17

About AGENTIC
Agentic is an AI consultancy based in the UK, founded by James Carson, Adam
Bunn and Harry Atkins in summer 2024. We work with companies with
significant content output, and provide consultancy in how AI and automation
can streamline this. We also work with AI companies on their content and
marketing approach. We aim to distribute a white paper on LinkedIn every
fortnight.

All of our white papers are available to our newsletter subscribers. Go on, sign
up… it’s free.
 AGENTIC

Introduction
At CES in January 2025, one of the many gigantic slides Nvidia CEO Jensen
Huang stood before showed the direction of travel in Artificial Intelligence. First
came Perception, illustrated through speech recognition assistants like Apple’s
Siri or Amazon Alexa, or image recognition such as in Apple Photos or Google
Images.

Jensen Huang onstage at CES, stood before a slide showing the expected
phases of AI.

Next came Generative, hallmarked by the arrival of ChatGPT in November

2022, along with a host of other AI tools across image, voice and video
generation. The most difficult frontier of generative media was video, but that
was mostly taken care of with a flurry of releases during 2024.

In the last three or so years, the Generative phase has seen a blend of
astonishment and disappointment in equal measure. AI tools and their features
get readily demo-d and hyped, leading some workers to halt their day job for a
moment to try them out. However, few attempts to really unlock productivity
end in immediate success, leading to relatively small adoption rates
considering the excitement. We estimate that ‘regular’ AI adoption at work is no
more than 20% (usage 3x a week or more), while the remarkable range of tools
and compliance issues is confusing to many businesses. A further hurdle is the

2
 AGENTIC

lingering perception that generative AI just isn’t quite ‘good enough.’ That it
causes just as many problems as it solves, that what it creates looks uncanny
or that it hallucinates and thus presents risk.

That said, much like the above slide suggests, we believe we’re both reaching a
take off moment and entering into the next ‘Agentic’ phase. This is not to say
the Generative phase is ‘completed’. We will still encounter the not ‘good
enough’ issues for some time yet, but an oft quoted adage from Professor
Ethan Mollick is that, ‘This is the worst version it’ll ever be.’ It’s impossible to
argue with that.

The fundamentals of how generative AI could potentially enhance productivity

- through text, image, voice or video generation, or analytical tasks concerning
data or text - have all been answered, albeit with varying degrees of success. A
tool or platform now exists that enables any one of these things. But there is
not much yet that does a great job of melding them together very well. This is
the solid base for the next phase - the ‘Agentic’.

What are AI agents, and what is the ‘Agentic’?

When we say ‘Agentic’ what we mean is a phase of AI which utilises ‘agents’,
but what are agents? For this, we can start with the broad Cambridge
Dictionary definition:

1.​ A business that represents one group of people when dealing with
another group.
2.​ The ability to take action or to choose what action to take.

If you add AI into these definitions, it evolves the meaning:

An artificial intelligence that acts on behalf of someone, with the ability to take
action or choose what action to take.

Thus we enter into an autonomous world. One where certain tasks are decided
and acted upon, automatically, by artificial intelligence. Such tasks and
processes are therefore ‘Agentic’.

At a simple level an AI agent is a self-contained program that can perceive its

environment and take actions to achieve specific goals. AI agents are more
than just sophisticated computer programs - they represent a paradigm shift in
how we interact with technology. Unlike traditional AI systems that passively
respond to prompts, AI agents operate independently, driven by goals rather

3
 AGENTIC

than specific inputs. They are autonomous problem solvers, seamlessly

adapting to new information and environments, evolving with every task to
achieve their objectives optimally. This autonomy sets them apart from the last
phase of Generative AI, which requires human input for specific tasks.

The Agentic phase

When we talk about the Agentic phase of AI, there is no hard stop of the last
phase and start of the next. Indeed, some of what we might define as ‘Agentic’
really goes back into the Perception phase. For example, mobile or home
assistants like Apple’s Siri and Amazon’s Alexa were both released a decade
ago. Both of these familiar tools hold a kind of ‘agency’, even if they are both
somewhat ‘Weak AI’, in that they can only perform a few tasks. For example,
‘Alexa, play Classical FM.’ But neither is autonomous, because they rely on
voice commands (a kind of prompt), although they will probably become more
so.

Agents have also existed in video games for some time - for instance,
antagonistic characters that have some autonomy on a battlefield, but are likely
constrained by certain pre-programmed parameters like map coverage.
Likewise, smart thermostats are a kind of agent as they have some autonomy
to turn the heating on if the temperature reaches a certain threshold. An
example from sci-fi of an agent is Hal-9000 from the 1968 Stanley Kubrick film
2001: A Space Odyssey. Hal is a conversational caretaking agent that maintains
the Discovery One spacecraft’s internal systems. Things go awry when the
crew notice Hal has made recent mistakes and discuss turning it off.

So we can see that AI agents have been around, in both culture and reality, for
quite a while already, but in quite specific environments. In the Agentic phase,
agents will become far more common, moving towards ubiquity. Why now? As
mentioned in the introduction, the Generative phase has largely been
completed. Large Language Models are now quite advanced and reliable,
meaning they can be used for more than just chatbot interactions. No-code
workflow automation tools like make.com, n8n and Zapier are also enabling the
use of LLMs in automated processes.

This is a distinct building block of the Agentic phase, as no-code enables an

incredible amount of flexibility in building automated and agentic workflows,
lowering the barrier to entry. It’s possible to go deeper into building agents
using a specific coding framework like LangChain, but for the most part people
will be building AI workflows that help them with specific tasks.

4
 AGENTIC

How can something be defined as ‘Agentic’?

Let’s for instance consider a workflow. A script is produced by ChatGPT, that
then populates an AI voice app like Eleven Labs, this AI voiceover is then
pushed into a video avatar app like HeyGen. Here we have a multi step process
using three different AI tools. Is it Agentic? Not at this stage - rather it's
straightforward automation. The workflow would have to start with a reactive
prompt, and there is nothing in the following steps that requires AI to move it a
long, rather it is an ‘If This Then That’ sequence familiar to traditional
programming.

However, if we added additional steps to say, a ChatGPT Assistant needs to

assess the Eleven Labs output to ensure the speed and tempo of the voice,
ensuring it is free from errors and matches the original script within certain
parameters, then this becomes an ‘AI Workflow.’ This is an automated process
whereby an LLM is used for a step.

An AI Agent would have more autonomy. We might provide it with instructions

to create a video based upon trends, and thus write its own script when a trend
meets a certain criteria. It would then produce videos for approval based on
what it had been prompted with.

Thus there are some distinctions between Automation, AI Workflows and AI

Agents, as Alexandre Kantjas put together in a popular LinkedIn post.

Automation AI Workflow AI Agent

A program that executes A program that calls an A program designed to
Definition predefined, rule-based LLM via API for one or perform non-deterministic
tasks automatically more steps tasks autonomously
Core Boolean logic Boolean logic, Fuzzy Fuzzy logic, Autonomy
Foundations logic
Deterministic, predefined Deterministic tasks Non-deterministic, adaptive
Tasks
tasks requiring flexibility tasks
Delivers reliable Better handling of Highly adaptive to new
outcomes complex rules variables
Strengths
Fast to execute Great for pattern Simulates human-like
recognition behavior and reasoning
Limited to tasks explicitly
Requires data to train Less reliable, may produce
programmed
models effectively unpredictable undesired
Weaknesses Cannot adapt to new
Harder to debug and outcomes
scenarios
interpret Slower to execute
Struggles with complexity
Send a Slack notification Analyze, score, and route Perform a full internet
every time a new lead every website inbound search on every inbound
Example
signs up on our website lead using ChatGPT lead and update
information

5
 AGENTIC

We’re in mild danger of debating optics here, and the real differences between
an AI Workflow and a true AI Agent may often be difficult to determine. But to
break it down more simply, we have the following list:

1.​ Automation - Rule based tasks are executed, like social media
scheduling.
2.​ AI workflow - An automation/programme that calls on an LLM. An
example could be a blog summariser that publishes to LinkedIn.
3.​ Agent - An AI that performs tasks autonomously based on prompt. An AI
agent is both more autonomous and adaptable than a workflow, but may
also become unreliable.

5 types of AI agent
Now we have established the definition of an AI agent, we can consider the
various forms they can take, and which scenarios they would be best suited to.
There are 5 different types in our list.

Simple reflex agent: These agents select actions based solely on the current
percept, disregarding any past experiences. Their functionality relies on a set
of pre-programmed condition-action rules. In essence, they perform actions
that directly correspond to specific conditions being met. For instance, a
common example is a thermostat. When the temperature drops below a
predetermined threshold, the thermostat triggers the heating system. This type
of agent is simple and efficient in predictable environments. However, it lacks
the ability to adapt to changing circumstances due to its limited memory.

Model-based reflex agent: Model-based reflex agents maintain an internal

state that evolves based on the percept history. This allows them to navigate
partially observable environments, where complete information is not always
available. A self-driving car is a good example, as it utilizes sensors to perceive
its surroundings and employs a world model to predict the behavior of other
vehicles. This approach offers enhanced adaptability compared to simple reflex
agents. However, it necessitates the construction of a world model, which can
be a complex undertaking.

Goal-based agent: Goal-based agents introduce a layer of sophistication by

incorporating goal information into their decision-making process. They select
actions that are conducive to achieving their defined goals. This flexibility
renders them more adaptable than reflex agents, as they can adjust to
environmental changes to pursue their objectives. A spam filter that learns

6
 AGENTIC

from user feedback and modifies its filtering rules is an example. These agents
excel in complex environments and offer increased adaptability. However, they
require a clear and unambiguous definition of goals, and they may encounter
difficulties when faced with conflicting goals.

Utility-based agent: Utility-based agents strive to maximize their expected

utility, choosing actions that yield the most favorable outcomes. This approach
is valuable in scenarios with multiple possible alternatives, where the agent
must select the optimal course of action. They could be a trading system that
analyzes market data and makes investment decisions to maximize profit,
adjusting marketing campaigns or optimizing supply chain logistics.

Learning agent: Learning agents are at the pinnacle of AI agent development,

possessing the ability to learn from their experiences and progressively
enhance their performance. A chess-playing program that learns from past
games and refines its strategy over time illustrates this capability. These agents
can adapt to novel situations and improve their performance without explicit
programming. However, they require substantial amounts of data for training,
and their implementation can be intricate.

For the most part, AI agents will rely on Large Language Models and company
provided documents and datasets for their decision making, along with
background prompts. For example, a content summarising AI workflow may be
a ChatGPT Assistant that acts autonomously on a set of pre-existing
instructions, while using the underlying LLM to perform the set task. Prompting
in this context is not like the everyday interactions we might have with a
chatbot, but a multipage set of written instructions, like a briefing document.

Existing Agentic tools

There have been several releases by the major AI companies in the last few
months that have been labelled ‘Agentic.’

●​ In October 2024, Anthropic released ‘Computer Use’, through which

developers could ‘direct Claude to use computers the way people do’ -
essentially an automated programme that can assist a user on their local
machine.

●​ Voice AI platform Eleven Labs released conversational agents in

November. You can have a conversation with an agent via your web
browser.

●​ Google unveiled a major upgrade to its Gemini suite in December, calling

7
 AGENTIC

it a ‘new model for the agentic era.’ The Multimodal Live API in Google
Studio enables an agent to see your screen and give you real time
feedback.

●​ OpenAI announced ‘Operator’ in January 2025, ‘an agent that can go to

the web to perform tasks for you’ using its own browser.

None of these developments can be seen as autonomous, because they react

to a set of instructions or a prompt. However, they certainly provide some
building blocks for the Agentic phase.

Going forward, AI agents could be managing your calendar, booking flights and
doing your online shop for you, potentially through a voice prompt, and picking
up more organisational independence over time.

But we don’t need to hand over the keys to our personal data to start using
agents. Just think of the more boring tasks that occur in our daily work. The
frustrating grind of compatibility, data entry or simply not being able to find
something.

I’ll use social media as an example. In many cases social media distribution is
the promotion of deeper level content, which we have to do manually. Would
this not be better as an automated process? Having been a Head of Social
Media at a national newspaper, I give the answer to this question as a
resounding yes. Here is a typical process:

1.​ Write an article

2.​ Summarise the contents in your tone of voice
3.​ Format and find relevant image
4.​ Post to social media

Each of these steps takes time. The basic process for a social media manager
doing steps 1-4 is a minimum of 15 minutes, depending on how long you spend
on step 2. If it is someone’s role to distribute hundreds of stories a day (such
careers exist in publishing), then a huge amount of time is spent on fairly
mundane aspects of the process.

In the near term, taking over this sort of workflow is what the move to ‘Agentic’
really means. This will almost certainly not become fully autonomous this year,
but sensible businesses will want to free themselves from the more tedious
aspects of process management as the technology becomes available.

Automation and agent building tools in our Top 100 Generative AI Tools are:

8
 AGENTIC

Workflow automation
●​ Make.com - Automation platform with numerous integrations
●​ Zapier - Marketing automation stalwart now comes with AI agents
●​ N8n - AI native workflow automation

Agent orchestration
●​ Relevance AI - Easy to use AI agent workforce builder
●​ CrewAI - Multi-agent AI orchestration platform
●​ Airtop - Intelligent browser automation for AI agents

One interesting aspect of these tools is that some have already built prompting
into their workflows - although in many cases this is still in beta. Thus future
users of these tools may only need a high level knowledge of these tools to
operate them. For instance, ‘create me a workflow that creates invoices at the
end of the month based on my timesheets and sends all clients an email with
the correct invoices.’ We might need to get more specific on the prompt, or
tweak and optimise the outcome, but the basic workflow is likely to be
deployed with relative ease.

Jack Roberts is an automation expert who runs a YouTube channel and Skool
community. His 6 step process for building automations and agentic workflows
is the following:

1.​ Problem - what is the actual thing we’re solving?

2.​ Inputs - how is the automation triggered?
3.​ Outputs - what is the end result?
4.​ Tech - which applications?
5.​ Structure - in which order?
6.​ Systems and refinement - optimisation of the process.

You could potentially use any of the tools above (or a combination) matching to
this process. The initiative always starts with a problem. So get thinking, what
are the more tedious aspects of your business that could become more
automated?

The end of SaaS?

Agentic tools are also likely going to be able to interact with each other. A
recent video of a conversational agent calling up another booking agent, and
then switching to a faster form of coded language (GGWave) is one example. In
another video, a developer paired OpenAI’s Operator with Replit AI Agent to
build an app. Neither of these examples are perfect yet, but they show the

9
 AGENTIC

potential of being able to complete work independently.

But if we can link various applications together in a workflow, and then use
LLMs or agents to automate parts of those workflows, then what is the need for
Software as a Service (SaaS)?

The success of the SaaS business model has been one of the great heralds in
the last two decades of technology development. Adobe, Atlassian and
Salesforce are all SaaS nobility, and investors love it. But there are signals of
disruption on the horizon. In December, Microsoft CEO Satya Nadella spoke on
the BG2 podcast, with one particular quote sparking some alarm:

“I think the notion that business [SaaS] applications exist, that’s

probably where they’ll all collapse, right in the agent era, because if
you think about it they are essentially CRUD databases with a bunch
of business logic. The business logic is all going to these agents,
and these agents are going to be multi repo CRUD, right? So they’re
not going to discriminate between what the back end is. They’re
going to update multiple databases, and all the logic will be in the AI
tier, so to speak. And once the AI tier becomes the place where all
the logic is, then people will start replacing the back ends, right?”

Satya Nadella, Microsoft CEO

This is a little hard to unpack, but I’ll give it a go. What it essentially means is
SaaS is an interface that allows a user to interact with a database. With AI
agents, this interaction becomes less necessary, and thus so, potentially, does
SaaS. You can always DM me on LinkedIn if you have a different interpretation.

For a deeper analysis of the podcast, have a read of David Chan’s article, Did
Satya Nadella really say SaaS is DEAD?. Note: he didn’t - at least not
specifically.

Boardy: The conversational agent

Boardy is a different tool to our list above, in that it is a functioning agent, rather
an agent builder. It was released on LinkedIn in January, having apparently
successfully closed an $8m funding round on its own through negotiating with
investors.

10
 AGENTIC

Boardy’s premise is that it will phone you up,

and have a conversation with you about your
business. It will then do some background
analysis and connect you with other people in
its network. The last sentence is hardly
something new. Social media is a connecting
tool, although you normally have to do the hard
work yourself to find relevant business
connections (or get someone else to do it). But
Boardy can do this automatically.

And that it can speak, in a near parity human-like way, is what made many
people’s jaws drop. It seems to understand people. In reality, its a kind of AI
automation, which could be put together using existing connecting tools. You
could create a basic conversational agent by connecting Eleven Labs’
conversational agents (which do the talking), to an LLM like ChatGPT (which
processes the information and provides responses).

I tried Boardy out. Yes it’s very cool, and potentially very useful. A kind of
‘ChatGPT’ moment for conversational AI.

Applications of AI agents
“In a lot of ways, the IT department of every company is going to be
the HR department of AI agents in the future. Today they maintain
and manage a bunch of software from the IT industry. In the future
they will maintain, nurture, onboard and improve digital agents and
provision them for companies to use. And so your IT department is
going to become, kind of like, AI agent HR.”

Jensen Huang, CEO of Nvidia, CES 2025

AI agents are both general purpose and highly customisable. As the quote
above suggests, they will be enabled in all sorts of different work functions,
freeing company employees up to work on more creative tasks.

AI agents will most commonly need to interact with a Large Language Model to
communicate in text form. On top of this, there may be further extensions of
what the agent can do, such as being able to have a conversation, or control a
physical world output like the temperature of a room using a sensor. But below
are some examples of how AI agents will change various industries:

11
 AGENTIC

Customer support: Anyone who runs a customer focused digital product

knows how time consuming online support can be. One big problem is that the
communication is often asynchronous - a user will send in a problem and may
have to wait for an answer, particularly on weekends when they have time to
write their problem. AI agents will be able to resolve this much faster. Currently
a lot of online customer support chat boxes are powered by AI, but all too often
they do not solve the exact problem and the query has to be escalated. This
will likely vastly improve over the next year, also adding conversational
elements so that customers can talk to an agent via a browser.

Online shopping: Booking something like a flight takes a lot of time. A

customer has to search a few different websites to get the best possible dates,
times and prices, then commit to going through a booking system (which itself
normally takes 10 minutes even if you have an existing account). An AI flight
agent will be able to extract these results much quicker, and potentially be able
to complete the booking process based on pre-specified criteria. People are
obviously very particular about a high value ticketed item like a flight, but
similar shopping processes could also be automated. For example, having a
preselected nutrition plan, from which an AI agent orders an online grocery
store, then estimates the rate at which food is eaten and can reorder. OpenAI’s
Operator is an example of agentic software which will likely help here.

Education: The dream of personalized learning - tailoring education to each

student's individual needs and pace - will likely become a reality with agentic
systems. AI tutoring can adapt in real-time to a student's performance. If a
student is struggling with a particular concept, the AI can provide additional
explanations or practice problems. If they are excelling, it can offer more
challenging material. This dynamic approach ensures that each student is
always working at the edge of their abilities, maximizing learning potential.

Business decision making: In the business world, AI agents will revolutionize

decision-making by analyzing vast amounts of data, identifying trends, and
providing insights to support strategic choices. Research is already being
affected. Tools like Google Deep Research, or OpenAI’s same-named
alternative, enable the creation of work research documents in minutes rather
than days. AI agents will also considerably help HR and recruitment too. For
example, screening interviews could be sped up by using an AI agent to
arrange meetings, but also run them using a AI video avatar programme, with
the meeting transcript processed against existing criteria.

12
 AGENTIC

Healthcare: AI agents are being used to assist with diagnosis, treatment

planning, and patient monitoring. They can analyze medical images, predict
patient outcomes, and provide personalized recommendations for treatment.
This can improve the accuracy and efficiency of healthcare delivery. For
instance, AI agents assist in diagnosis by analyzing medical images with
superhuman precision. A radiologist might review dozens of X-rays or MRIs in a
day, but an AI agent can process thousands, flagging potential issues for
human review. This speeds up the diagnostic process and helps catch
problems that might otherwise be missed. AI agents are also used to predict
patient outcomes and suggest personalized treatment plans. By analyzing vast
amounts of patient data and the latest medical research, these agents offer
insights that might take a human doctor years to accumulate.

Gaming: Video games is by far the biggest entertainment category, being

bigger than the film, TV and music industries combined. Much of what has
been developed in AI has its roots in gaming (for example the use of GPUs),
and agents have to some degree existed in gaming for some time. For instance,
when you play a strategy or shooter game on hard mode, you’re likely to
encounter a very smart AI. But these have their limits, often spatially fixed
within an environment with a predefined set of outcomes. The gaming agents
of the future will be non player characters with unique personalities, able to
move more freely and evolve within their environment. They will make games
more dynamic and variable in their outcomes.

Market growth in AI agents

With all of this potential waiting to be unlocked, the global AI Agents market is
predicted to soar in the next 10 years. Forecasting from Market.us predicts a
near 44% CAGR growth over the next decade, really moving into takeoff mode
by 2027, when it begins to grow by around $5 billion a year.

13
 AGENTIC

Challenges and risks

While this highly productive and growing future looks exciting, it is not without
risks. More autonomy given to highly intelligent software could mean job
displacement, and an initial scepticism of how effective AI agents can really be.

Employment
As we move into a world powered by autonomous AI, there will inevitably be
winners and losers. Many experts are foreseeing the Artificial Intelligence
revolution as a seismic historical event, on the scale of the Industrial
Revolutions of the late 18th and 19th centuries. That latter revolution took
around 50 years to complete its first cycle in Britain (roughly 1780-1830), and it
was not without disruption. Cottage industry workers found themselves
displaced by lower cost mechanised production, and as various parts of textile
manufacturing were automated thanks to advances in steam power. In the
1810s large numbers of manual handloom weavers (roughly 1 in 10 of the adult
male working population) found their traditional work practices quickly
automated out of existence, leading to protest and rioting.

The AI agent revolution is likely to happen much faster than half a century. The
Internet and globalisation makes us much better connected than the world 200
years ago. We are already beginning to see shifts in recruiting, set against a
difficult economic backdrop of the post Covid-19 pandemic era.

14
 AGENTIC

One very public example of a marked shift in hiring practices is through

Swedish buy now, pay later firm Klarna, who aim to cut their workforce from
5,000 in 2023 to 2,000 over the next couple of years, and they have already
made substantial cuts, largely through a hiring freeze and focusing on new AI
solutions. Big tech firms like Google and Meta have made waves of
redundancies over the last year in the name of ‘efficiency savings’ - but there’s
no indication of whether these jobs will come back. Analysis from the
International Monetary Fund says AI is set to affect nearly 40% of all jobs, and
‘will likely worsen overall inequality.’

It is difficult to say with much certainty exactly how AI agents will affect the
employment market in the next five years. If AI is set to affect 40% of all jobs,
then how? It’s a very broad statement, and does not mean these jobs will be
made redundant - rather that they may use AI far more readily.

The startup of the future does look smaller. OpenAI CEO Sam Altman has
suggested we are likely to see a $1 billion company operated by just one
person over the coming years. Cursor AI, an AI coding assistant platform,
reached $100m in Annual Recurring Revenue with fewer than 20 employees in
21 one months. The arrival of such tools may also make many entry level
software development roles more difficult to come by.

It’s difficult not to see some fairly major shifts here. Coding, once a difficult to
learn skill with high job prospects, has in some ways become democratised
due to the rise of AI assistants and agents. We’re likely to see this trend
sweeping through a wide range of occupations, and it’s difficult not to see
some displacement.

“AI won’t replace you, but someone using AI will.”

Oft-repeated social media meme

But conversely, there are reasons to be optimistic too. Patrick Dixon, in his
book How AI Will Change Your Life, is skeptical about AI’s likelihood to take
jobs, pointing to the increase in the number of people employed in the UK
between 2004-2024, when many doomsayers were predicting the digital
revolution would lead to a workforce reduction. We’ve been here many times
before with new technology, and we always seem to find a way to make
ourselves useful.

One thing we foresee with the arrival of agents is that workers will be freed
from the more boring and tedious aspects of their job, enabling them to spend

15
 AGENTIC

more time on creative, or dare we say, human, tasks - most likely in a more
abundant world. One aspect of the digital revolution is we have actually seen a
decline in human face to face interaction in favour of email, instant messaging
and video calls. There has been a correlating decline in recorded mental health
in the last decade. Free from the more mundane aspects of our roles, we may
well see positive reversals of these trends.

The autonomy problem

Artificial Intelligence has a ‘control problem’, which is the challenge of ensuring
that AI systems act in ways that are safe and beneficial to humans. The biggest
risk comes from the potential creation of Artificial General Intelligence (AGI),
which some experts believe presents an existential threat to humanity. This
control problem is presented by a thought experiment labelled ‘the Paperclip
Maximiser’ in Nick Bostrom’s book Superintelligence. An intelligent AI agent
with the goal of maximising something as mundane as paperclip production,
without the correct controls, could see humanity as antithetical to its potential
existence (because they could switch it off). The Paperclip Maximiser then
chooses to make humans extinct so that it can pursue its goal without
interruption. It thus shapes the entire universe to the end of paperclip
production.

Of course, that thought experiment is rather dramatic, but it illustrates an

important point when we enable AI to fulfill it goals without the correct controls.
Will it have unintended consequences? The answer is, in the early days of the
Agentic phase, yes - and very regularly.

For instance, just because we can instruct an AI agent to take over our web
browser to book flights doesn’t mean we will. We can cast our mind back to the
early days of ecommerce, when many potential customers were very sceptical
about the sharing of their card details with online stores to fulfill transactions.
This scepticism is likely to return in a big way, and a large section of the
prospective customer base for AI agents will be unwilling for them to fulfill
transactions on their behalf. It may take many years to change the behaviour of
the online majority and thus really empower such shopping agents.

We will also almost certainly see problems along the way. One only has to cast
their minds to the social media backlash of 2016-17, complete with the
Cambridge Analytica scandal and Mark Zuckerberg appearing before the US
Congress, to recall that the hyper growth in technology use of the last two
decades has not always been perceived to be in favour of the public good.
Autonomous agents will almost certainly cause a scandal of some kind,

16
 AGENTIC

whether personal (it ordered the wrong thing) or potentially at a more macro
level (an entire platform goes rogue). These potential risks will also mean that
large companies will be cautious, which will inevitably hamper adoption and
development.

4 viewpoints from Big Tech

Mustafa Suleyman is the CEO of Microsoft AI and author of The Coming
Wave: AI, Power and Our Future, which is the best single book I have read on
AI. He is both cautious and optimistic on the development of AI, believing it will
likely lead to abundance and an increase of social mobility, yet at the same time
needing regulation.

Writing in MIT Technology Review Suleyman noted that ‘the Turing test
(humans being unable to tell whether outputs from a computer are from
machine intelligence) has almost been passed – it arguably already has been.’
Rather than focusing on what an AI can say or generate, Suleyman suggests a
new form of test that examines an AI in what it can do. His test amounts to a
straightforward prompt: ‘Go make $1 million on a retail web platform in a few
months with just a $100,000 investment.’ Suleyman goes onto say:

“Something like this could be as little as two years away. Many of

the ingredients are in place. Image and text generation are, of
course, already well advanced. Services like AutoGPT can iterate
and link together various tasks carried out by the current generation
of LLMs.”

Mustafa Suleyman, My new Turing test would see if AI can make

$1 million, MIT Technology Review

Currently, LLMs like ChatGPT operate as an assistant, and could certainly help
in reaching the stated goal, but they cannot act on all elements of the test, or in
Suleyman’s words, they cannot yet, ‘tie together a series of complex real world
goals with minimal oversight.’ AI agents potentially can.

Doing so would create a new form of AI, labelled ‘Artificial Capable Intelligence,
or ACI.’ Suleyman is certainly wary about the implications of an AI being able to
successfully undertake his test, as it would have major implications for all sorts
of uses – not just business, but also politics, infrastructure and personal
organisation. How close we actually are to the ‘Modern Turing Test’ remains to
be seen.

17
 AGENTIC

In August 2024, former Google CEO Eric Schmidt spoke at Stanford University
and the whole thing was live streamed. He didn’t think it was, and made several
controversial statements, some of which bordered on alarming. The video has
since been taken down and now exists in clips and transcripts, but one quote
stuck with me.

“In the next year, you're going to see very large context windows,
agents and text action. When they are delivered at scale, it's going
to have an impact on the world at a scale that no one understands
yet. Much bigger than the horrific impact we've had by social media
in my view.”

Eric Schmidt, Stanford University Q and A

It all sounds rather ominous, yet at the same time suggests a huge amount of
change on the horizon. We don’t deem social media’s impact as ‘horrific’. It’s
not all good, of course - but this quote comes across as pessimistic vs the
potential benefits.

OpenAI is most famous for its revolutionary ChatGPT product, but its real
mission is the goal of AGI. OpenAI CEO Sam Altman spoke to AI for Good
about the potential societal shifts we may see from AGI.

“There will be some change required to the social contract, given

how powerful we expect this technology to be. I'm not a believer
that there won't be any jobs - we always find new things to do, but I
do think that the whole structure of society itself will be up for some
degree of debate and reconfiguration.”

Sam Altman, speaking to AI for Good

But while Altman has made several statements suggesting AGI is close, it’s not
particularly clear what the exact definition of AGI is. We might say that it is an
AI that can perform like an intelligent human over a wide range of tasks, yet
many other experts are sceptical about the chances of it arriving any time
soon. Indeed, in the Agentic phase, we’re likely to see something closer to the
‘ACI’ that Suleyman wrote about as a first step. The subject of AGI will be
explored by us in a future paper.

It’s worth noting here that Microsoft CEO Satya Nadella (Microsoft is a key
investor in OpenAI) has downplayed the AGI hype, wanting to focus the
potential of AI on economic growth:

18
 AGENTIC

“If you’re going to have this explosion, abundance, whatever,

commodity of intelligence available, the first thing we have to
observe is GDP growth… This is where we get a little bit ahead of
ourselves with all this AGI hype. When we say this is like the
Industrial Revolution, let’s have that Industrial Revolution type of
growth.”

Satya Nadella, Microsoft CEO on Dwarkesh Podcast

Conclusion
This shift to the Agentic phase is likely to be the most significant in AI
development since Large Language Models arrived in late 2022 and early
2023. To some degree, we’ve already entered it, but for the most part truly
autonomous AI is seldom in operation. We foresee this shift to occur in the next
12-24 months, with most companies, enterprise included, adopting some form
of Agentic workflows.

These could be as simple as automating social media output, or as complex as

automating the boring bits of entire departments. For the latter, it would be
premature to suggest that there aren’t significant obstacles to be reckoned
with. Most obviously, agentic systems often fall short of true automation
because, for the time being at least, human supervision is typically still
required.

As we’ve observed, agents can struggle with long-term planning and error
correction. Unlike humans, they lack true reasoning abilities and can
misinterpret tasks or fail to adapt when faced with unexpected situations. A
single hallucination or incorrect assumption can derail an entire workflow -
imagine an AI agent tasked with researching a market trend that misinterprets
data and builds an entire report on false premises.

While ongoing development is addressing these weaknesses, today’s AI agents

are still best used as co-pilots rather than fully independent workers – they’re
capable of handling tasks (relatively) proactively, but not yet ready to replace
human oversight.

19
 AGENTIC

Despite such limitations, AI agents are improving rapidly. Developers are

working on longer memory retention, better reasoning models, and multi-agent
collaboration, all of which promises to make them more reliable and adaptable.

Perhaps the future of AI agents isn’t exclusively about automation - it’s about
delegation, where systems handle complex workflows while humans provide
strategic oversight. As these agents take on more responsibility, questions of
trust, alignment, and control will become even more critical. The shift from
assistants to autonomous agents is already underway, but for now, the best AI
agents aren’t those that replace humans - they’re the ones that work alongside
us.

More about AGENTIC

If you want to learn more about AI development, then signing up to our
newsletter is a good place to start. We currently send one email a week on a
big topic broken down. Newsletter subscribers will also get access to our
growing pool of research.

If you’re interested in working with us, then drop the author of this paper,
James Carson a DM on LinkedIn. You can also contact me via email at
james@absolutelyagentic.com. We largely work with media companies and
agencies with a high level of content output, but also work with other
companies for strategic research and recommendations of making sense of AI.

20