McKinsey Digital Practice

Why agents are the next

frontier of generative AI
By moving from information to action—think virtual coworkers
able to complete complex workflows—the technology promises a
new wave of productivity and innovation.
By Lareina Yee, Michael Chui, and Roger Roberts
with Stephen Xu

July 2024
Over the past couple of years, the world has agent system to accomplish a complex workflow.
marveled at the capabilities and possibilities A multiagent system could then interpret and
unleashed by generative AI (gen AI). Foundation organize this workflow into actionable tasks, assign
models such as large language models (LLMs) can work to specialized agents, execute these refined
perform impressive feats, extracting insights and tasks using a digital ecosystem of tools, and
generating content across numerous mediums, collaborate with other agents and humans to
such as text, audio, images, and video. But the next iteratively improve the quality of its actions.
stage of gen AI is likely to be more transformative.
In this article, we explore the opportunities that
We are beginning an evolution from the use of gen AI agents presents. Although the
knowledge-based, gen AI–powered technology remains in its nascent phase and
tools—say, chatbots that answer questions requires further technical development before
and generate content—to gen AI–enabled “agents” it’s ready for business deployment, it’s quickly
that use foundation models to execute complex, attracting attention. In the past year alone, Google,
multistep workflows across a digital world. In short, Microsoft, OpenAI, and others have invested in
the technology is moving from thought to action. software libraries and frameworks to support
agentic functionality. LLM-powered applications
Broadly speaking, “agentic” systems refer to such as Microsoft Copilot, Amazon Q, and Google’s
digital systems that can independently interact upcoming Project Astra are shifting from being
in a dynamic world. While versions of these knowledge-based to becoming more action-based.
software systems have existed for years, the Companies and research labs such as Adept,
natural-language capabilities of gen AI unveil new crewAI, and Imbue also are developing agent-based
possibilities, enabling systems that can plan their models and multiagent systems. Given the speed
actions, use online tools to complete those tasks, with which gen AI is developing, agents could
collaborate with other agents and people, and learn become as commonplace as chatbots are today.
to improve their performance. Gen AI agents
eventually could act as skilled virtual coworkers,
working with humans in a seamless and natural What value can agents bring
manner. A virtual assistant, for example, could plan to businesses?
and book a complex personalized travel itinerary, The value that agents can unlock comes from their
handling logistics across multiple travel platforms. potential to automate a long tail of complex use
Using everyday language, an engineer could cases characterized by highly variable inputs and
describe a new software feature to a programmer outputs—use cases that have historically been
agent, which would then code, test, iterate, and difficult to address in a cost- or time-efficient
deploy the tool it helped create. manner. Something as simple as a business trip, for
example, can involve numerous possible itineraries
Agentic systems traditionally have been difficult encompassing different airlines and flights, not
to implement, requiring laborious, rule-based to mention hotel rewards programs, restaurant
programming or highly specific training of reservations, and off-hours activities, all of which
machine-learning models. Gen AI changes that. must be handled across different online platforms.
When agentic systems are built using foundation While there have been efforts to automate parts of
models (which have been trained on extremely large this process, much of it still must be done manually.
and varied unstructured data sets) rather than This is in large part because the wide variation in
predefined rules, they have the potential to adapt potential inputs and outputs makes the process too
to different scenarios in the same way that LLMs complicated, costly, or time-intensive to automate.
can respond intelligibly to prompts on which they
have not been explicitly trained. Furthermore, using Gen AI–enabled agents can ease the automation of
natural language rather than programming code, complex and open-ended use cases in three
a human user could direct a gen AI–enabled important ways:

Why agents are the next frontier of generative AI 2

— Agents can manage multiplicity. Many business defining characteristic of agents (it’s one way
use cases and processes are characterized by that they can act in the world) but also a way in
a linear workflow, with a clear beginning and which their gen AI capabilities can uniquely be
series of steps that lead to a specific resolution brought to bear. Foundation models can learn
or outcome. This relative simplicity makes them how to interface with tools, whether through
easily codified and automated in rule-based natural language or other interfaces. Without
systems. But rule-based systems often exhibit foundation models, these capabilities would
“brittleness”—that is, they break down when require extensive manual efforts to integrate
faced with situations not contemplated by the systems (for example, using extract, transform,
designers of the explicit rules. Many workflows, and load tools) or tedious manual efforts to
for example, are far less predictable, marked by collate outputs from different software systems.
unexpected twists and turns and a range of
possible outcomes; these workflows require
special handling and nuanced judgment that How gen AI–enabled agents could work
makes rules-based automation challenging. Agents can support high-complexity use cases
But gen AI agent systems, because they are across industries and business functions,
based on foundation models, have the potential particularly for workflows involving time-consuming
to handle a wide variety of less-likely situations tasks or requiring various specialized types of
for a given use case, adapting in real time to qualitative and quantitative analysis. Agents do this
perform the specialized tasks required to bring by recursively breaking down complex workflows
a process to completion. and performing subtasks across specialized
instructions and data sources to reach the desired
— Agent systems can be directed with natural goal. The process generally follows these four
language. Currently, to automate a use case, it steps (Exhibit 1):
first must be broken down into a series of rules
and steps that can be codified. These steps are 1. User provides instruction: A user interacts with
typically translated into computer code and the AI system by giving a natural-language
integrated into software systems—an often prompt, much like one would instruct a trusted
costly and laborious process that requires employee. The system identifies the intended
significant technical expertise. Because agentic use case, asking the user for additional
systems use natural language as a form of clarification when required.
instruction, even complex workflows can
be encoded more quickly and easily. What’s 2. Agent system plans, allocates, and executes
more, the process can potentially be done by work: The agent system processes the prompt
nontechnical employees, rather than software into a workflow, breaking it down into tasks and
engineers. This makes it easier to integrate subtasks, which a manager subagent assigns to
subject matter expertise, grants wider access other specialized subagents. These subagents,
to gen AI and AI tools, and eases collaboration equipped with necessary domain knowledge
between technical and nontechnical teams. and tools, draw on prior “experiences” and
codified domain expertise, coordinating with
— Agents can work with existing software tools each other and using organizational data and
and platforms. In addition to analyzing and systems to execute these assignments.
generating knowledge, agent systems can use
tools and communicate across a broader digital 3. Agent system iteratively improves output:
ecosystem. For instance, an agent can be Throughout the process, the agent may request
directed to work with software applications additional user input to ensure accuracy and
(such as plotting and charting tools), search relevance. The process may conclude with the
the web for information, collect and compile agent providing final output to the user, iterating
human feedback, and even leverage additional on any feedback shared by the user.
foundation models. Digital-tool use is both a

Why agents are the next frontier of generative AI 3

Web 2024
McKQ-2024Q4-GenAIAndSmartAgents
Exhibit 1 of 2
Exhibit 1

Agents enabled by generative AI soon could function as hyperefficient

virtual coworkers.

Illustration of how an Agent system

agent system might
execute a workflow, Manager agent
from prompt to output

External systems:
Agents interact with
databases and systems—
both organizational and
Analyst Checker Planner external data—to
agent agent agent complete the task.

Specialist agents

Start 1 2 3 4 End

Using natural language, The agent system interprets The agent team shares The agent team receives user
the user prompts the the prompt and builds a the draft output with feedback, then iterates and
generative AI agent work plan. A manager agent the user. refines output accordingly.
system to complete a subdivides the project into
task. tasks assigned to specialist
agents; they gather and
analyze data from multiple
sources and collaborate
with one another to execute
their individual missions.

McKinsey & Company

4. Agent executes action: The agent executes any required, this tends to be a time-consuming and
necessary actions in the world to fully complete highly collaborative effort, requiring a relationship
the user-requested task. manager to work with the borrower, stakeholders,
and credit analysts to conduct specialized analyses,
which are then submitted to a credit manager for
Art of the possible: Three potential review and additional expertise.
use cases
What do these kinds of systems mean for Potential agent-based solution: An agentic
businesses? The following three hypothetical system—comprising multiple agents, each assuming
use cases offer a glimpse of what could be possible a specialized, task-based role—could potentially
in the not-too-distant future. be designed to handle a wide range of credit-risk
scenarios. A human user would initiate the process
Use case 1: Loan underwriting by using natural language to provide a high-level
Financial institutions prepare credit-risk memos to work plan of tasks with specific rules, standards, and
assess the risks of extending credit or a loan to a conditions. Then this team of agents would break
borrower. The process involves compiling, analyzing, down the work into executable subtasks.
and reviewing various forms of information pertaining
to the borrower, loan type, and other factors. Given One agent, for example, could act as the
the multiplicity of credit-risk scenarios and analyses relationship manager to handle communications

Why agents are the next frontier of generative AI 4

between the borrower and financial institutions. Use case 2: Code documentation and
An executor agent could compile the necessary modernization
documents and forward them to a financial analyst Legacy software applications and systems at
agent that would, say, examine debt from cash flow large enterprises often pose security risks and
statements and calculate relevant financial ratios, can slow the pace of business innovation. But
which would then be reviewed by a critic agent to modernizing these systems can be complex, costly,
identify discrepancies and errors and provide and time-intensive, requiring engineers to review
feedback. This process of breakdown, analysis, and understand millions of lines of the older
refinement, and review would be repeated until codebase and manual documentation of business
the final credit memo is completed (Exhibit 2). logic, and then translating this logic to an updated
codebase and integrating it with other systems.
Unlike simpler gen AI architectures, agents can
produce high-quality content, reducing review cycle Potential agent-based solution: AI agents have
times by 20 to 60 percent. Agents are also able to the potential to significantly streamline this process.
traverse multiple systems and make sense of data A specialized agent could be deployed as a
pulled from multiple sources. Finally, agents can legacy-software expert, analyzing old code and
show their work: credit analysts can quickly drill documenting and translating various code segments.
into any generated text or numbers, accessing the Concurrently, a quality assurance agent could
complete chain of tasks and using data sources to critique this documentation and produce test cases,
produce the generated insights. This facilitates the helping the AI system to iteratively refine its output
rapid verification of outputs. and ensure its accuracy and adherence to

Web 2024
McKQ-2024Q4-GenAIAndSmartAgents
Exhibit 2
Exhibit 2 of 2

Generative AI agents have the potential to change the way we work by

supercharging productivity.

Illustrative use case: credit-risk memos

Financial institutions A relationship The RM and a The credit analyst The RM reviews the memo
often spend 1–4 manager (RM) credit analyst typically spends and provides feedback;
weeks creating a gathers data from collaboratively 20+ hours writing the credit analyst writes
credit-risk memo. 15+ sources on analyze the data. the memo. a new draft incorporating
The current process: borrower, loan type, the feedback.
and other factors.

Start 1 2 3 4 End

Start 1 2 3 End

Generative AI The RM prompts The agent subdivides The RM and credit

(gen AI) agents could the gen AI agent the project into tasks analyst review the
cut time spent on system and that are assigned to memo and give
creating credit-risk provides relevant specialist agents, which feedback; the agent
memos by 20–60% materials needed gather and analyze data incorporates the
using these steps: to produce the from multiple sources feedback into the
memo. and then collaborate to final memo.
generate a draft memo.

McKinsey & Company

Why agents are the next frontier of generative AI 5

organizational standards. The repeatable nature of such, it is not too soon for business leaders to learn
this process, meanwhile, could produce a flywheel more about agents and consider whether some of
effect, in which components of the agent framework their core processes or business imperatives can be
are reused for other software migrations across the accelerated with agentic systems and capabilities.
organization, significantly improving productivity and This understanding can inform future road map
reducing the overall cost in software development. planning or scenarios and help leaders stay at the
edge of innovation readiness. Once those potential
Use case 3: Online marketing campaign creation use cases have been identified, organizations can
Designing, launching, and running an online begin exploring the growing agent landscape,
marketing campaign tends to involve an array utilizing APIs, tool kits, and libraries (for example,
of different software tools, applications, and Microsoft Autogen, Hugging Face, and LangChain)
platforms. And the workflow for an online to start understanding what is relevant.
marketing campaign is highly complex. Business
objectives and market trends must be translated To prepare for the advent of agentic systems,
into creative campaign ideas. Written and visual organizations should consider these three factors,
material must be created and customized for which will be key if such systems are to deliver on
different segments and geographies. Campaigns their potential:
must be tested with user groups across various
platforms. To accomplish these tasks, marketing — Codification of relevant knowledge:
teams often use different forms of software and Implementing complex use cases will likely
must move outputs from one tool to another, which require organizations to define and document
is often tedious and time-consuming. business processes into codified workflows that
are then used to train agents. Likewise,
Potential agent-based solution: Agents can help organizations might consider how they can
connect this digital marketing ecosystem. For capture subject matter expertise, which will be
example, a marketer could describe targeted used to instruct agents in natural language, thus
users, initial ideas, intended channels, and other streamlining complex processes.
parameters in natural language. Then, an agent
system—with assistance from marketing — Strategic tech planning: Organizations will
professionals—would help develop, test, and iterate need to organize their data and IT systems to
different campaign ideas. A digital marketing ensure that agent systems can interface
strategy agent could tap online surveys, analytics effectively with existing infrastructure. That
from customer relationship management solutions, includes capturing user interactions for
and other market research platforms aimed at continuous feedback and creating the flexibility
gathering insights to craft strategies using to integrate future technologies without
multimodal foundation models. Agents for content disrupting existing operations.
marketing, copywriting, and design could then build
tailored content, which a human evaluator would — Human-in-the-loop control mechanisms: As
review for brand alignment. These agents would gen AI agents begin interacting with the real world,
collaborate to iterate and refine outputs and align control mechanisms are essential to balance
toward an approach that optimizes the campaign’s autonomy and risk (see sidebar, “Understanding
impact while minimizing brand risk. the unique risks posed by agentic systems”).
Humans must validate outputs for accuracy,
compliance, and fairness; work with subject
How should business leaders prepare matter experts to maintain and scale agent
for the age of agents? systems; and create a learning flywheel for
Although agent technology is quite nascent, ongoing improvement.Organizations should start
increasing investments in these tools could result in considering under what conditions and how
agentic systems achieving notable milestones and such human-in-the-loop mechanisms should
being deployed at scale over the next few years. As be deployed.

Why agents are the next frontier of generative AI 6

Understanding the unique risks posed by agentic systems

Large language models (LLMs), as we Misuse of tools is key to ensuring that users maintain
now know, are prone to mistakes and With their ability to access tools and data, a balanced perspective, critically
hallucinations. Because agent systems agents could be dangerous if intentionally evaluate agent performance, and retain
process sequences of LLM-derived misused. Agents, for example, could be final authority and accountability in
outputs, a hallucination within one of used to develop vulnerable code, create agent actions. Furthermore, agent
these outputs could have cascading convincing phishing scams, or hack performance should be evaluated by tying
effects if protections are not in place. sensitive information. agents’ activities to concrete outcomes
Additionally, because agent systems (for example, customer satisfaction,
Mitigation strategy: For potentially
are designed to operate with autonomy, successful completion rates of tickets).
high-risk scenarios, organizations should
business leaders must consider additional
build in guardrails (for example, access In addition to addressing these potential
oversight mechanisms and guardrails.
controls, limits on agent actions) and risks, organizations should consider the
While it is difficult to fully anticipate all the
create closed environments for agents broader issues raised by gen AI agents:
risks that will be introduced with agents,
(for instance, limit the agent’s access
here are some that should be considered. — Value alignment: Because agents are
to certain tools and data sources).
akin to coworkers, their actions should
Potentially harmful outputs Additionally, organizations should apply
embody organizational values. What
Large language models are not always real-time monitoring of agent activities
values should agents embody in their
accurate, sometimes providing incorrect with automated alerts for suspicious
decisions? How can agents be
information or performing actions with behavior. Regular audits and compliance
regularly evaluated and trained to
undesirable consequences. These risks checks can ensure that guardrails remain
align with those values?
are heightened as generative AI (gen AI) effective and relevant.
agents independently carry out tasks — Workforce shifts: By completing tasks
Insufficient or excessive human–agent
using digital tools and data in highly independently, agent systems stand
trust
variable scenarios. For instance, an agent to significantly alter the way work is
Just as in relationships with human
might approve a high-risk loan, leading to accomplished, potentially allowing
coworkers, interactions between humans
financial loss, or it may make an expensive, humans to focus more on higher-level
and AI agents are based on trust. If users
nonrefundable purchase for a customer. tasks that require critical thinking and
lack faith in agentic systems, they might
managerial skills. How will roles and
Mitigation strategy: Organizations scale back the human–agent interactions
responsibilities shift in each business
should implement robust accountability and information sharing that agentic
function? How can employees be
measures, clearly defining the systems require if they are to learn and
provided with retraining
responsibilities of both agents and improve. Conversely, as agents become
opportunities? Are there new
humans while ensuring that agent outputs more adept at emulating humanlike
collaboration models that can
can be explained and understood. This behavior, some users could place too
enhance cooperation between
could be accomplished by developing much trust in them, ascribing to them
humans and AI agents?
frameworks to manage agent autonomy human-level understanding and judgment.
(for example, limiting agent actions This can lead to users uncritically — Anthropomorphism: As agents
based on use case complexity) and accepting recommendations or giving increasingly have humanlike
ensuring human oversight (for example, agents too much autonomy without capabilities, users could develop an
verifying agent outputs before execution sufficient oversight. overreliance on them or mistakenly
and conducting regular audits of agent believe that AI assistants are fully
Mitigation strategy: Organizations can
decisions). Additionally, transparency and aligned with their own interests and
manage these issues by prioritizing
traceability mechanisms can help users values. To what extent should
the transparency of agent decision
understand the agent’s decision making humanlike characteristics be
making, ensuring that users are trained
process to identify potentially fraught incorporated into the design of agents?
in the responsible use of agents, and
issues early. What processes can be created to
establishing a humans-in-the-loop
enable real-time detection of potential
process to manage agent behavior.
harms in human–agent interactions?
Human oversight of agent processes

Why agents are the next frontier of generative AI 7

 McKinsey’s most recent “State of AI” survey found require considerable testing, training, and
Find more content like this on the
that more than 72 percent of companies surveyed coaching before they can be trusted to operate
McKinsey Insights App
are deploying AI solutions, with a growing interest in independently. But even in these earliest of days,
gen AI. Given that activity, it would not be surprising it’s not hard to envision the expansive opportunities
to see companies begin to incorporate frontier this new generation of virtual colleagues could
technologies such as agents into their planning potentially unleash.
processes and future AI road maps. Agent-driven
automation remains an exciting proposition, with
the potential to revolutionize whole industries,
bringing a new speed of action to work. We are celebrating the 60th birthday of the
Scan • Download • Personalize
McKinsey Quarterly with a yearlong campaign
That said, the technology is still in its early stages, featuring four issues on major themes related to the
and there is much development required before its future of business and society, as well as related
full capabilities can be realized. The increased interactives, collections from the magazine’s
complexity and autonomy of these systems pose archives, and more. This article is part of the
a host of challenges and risks. And if deploying AI campaign’s Future of Technology issue. Sign up for
agents is akin to adding new workers to the team, the McKinsey Quarterly alert list to be notified as
just like their human team members, agents will soon as other new Quarterly articles are published.

Lareina Yee is a senior partner in McKinsey’s Bay Area office, where Michael Chui and Roger Roberts are partners;
Stephen Xu is a senior director of project management in the Toronto office.

The authors wish to thank Aneri Shah, Arun Mittal, Henry Zhang, Kimberly Te, Mara Pometti, and Rickard Ström for their
contributions to this article.

Designed by McKinsey Global Publishing

Copyright © 2024 McKinsey & Company. All rights reserved.

Why agents are the next frontier of generative AI 8