I have seen things I wouldn’t have believed even a few years ago. ChatGPT drafting content strategies from a three-sentence prompt. Grammarly solving my Oxford comma woes across an entire manuscript. I have yet to watch C-beams glitter in the dark. But I’ve witnessed AI reshape how I work — and it’s only just begun.

One area I find most compelling is agentic AI. Right now, AI agents sit squarely in the “next generation” of AI tools: developing quickly but not quite ready for the limelight. Still, Deloitte’s latest State of Generative AI in the Enterprise report urges companies to prepare their strategies and workflows for agentic AI.

Download Now: The Annual State of Artificial Intelligence in 2024 [Free Report]

You should know a thing or two about AI agents and how they can drive growth through AI workflow automation. Let’s investigate agentic AI and see how its potential could affect your company in the future.

Table of Contents

Agentic AI differs from the larger conversation happening around AI. Most workplace AI tools are “assistive AI” like Grammarly or “generative AI” like ChatGPT.

They have amazing capabilities but still require direct user input to operate (i.e., I need to enter a prompt into ChatGPT to make it work). Agentic AI can respond to user inputs but also can proactively pursue objectives, adjust to feedback, and run with some degree of self-sufficiency.

Notably, AI agents can run multi-step workflows automatically and adapt their processes in real time through feedback and new data. That’s a lot of power to grant a non-human operator within a business environment. As such, agentic AI does not make humans obsolete.

Instead, I believe human oversight of agentic AI will be necessary to deploy these tools wisely and ethically.

How do AI agents work?

how do ai agents work?

An AI agent overcomes traditional AI’s limitations to enable problem-solving, decision-making, and influence over external environments. While they can automate lower-level, repetitive tasks, they really excel at adapting to dynamic environments and optimizing outcomes over time.

But how do they actually accomplish that? The short version: agentic AI operates with a few key steps differing from other AI systems you might’ve tried before.

Let’s say you give an AI agent a task like, “Schedule a recurring weekly meeting with the five members of my marketing team.” How would agentic AI complete this request?

1. Agents define the goal and task steps.

The AI agent begins by processing the objective — in this case, scheduling a recurring meeting with specific people on a certain time frame. Some agents can develop this objective autonomously based on context, an important feature in multi-agent operations.

For now, though, this agent will work with the human-based request.

Behind the chat window, the AI agent uses Natural Language Understanding (NLU) to interpret the prompt and pull out key details. Then, it’ll deploy a combination of reasoning models like a Large Language Model (LLM) to understand context and structured task planners to divide the objective into smaller operational subtasks.

For our example, the agent might build a list like:

  • Gather the team’s availability.
  • Identify date and time conflicts.
  • Find the optimal time for the entire team.
  • Send meeting invites and follow-up messages.

This gives the machine specific next steps to develop instructions for its own operation.

2. Agents plan and reason through multiple steps.

The AI agent won’t just grab the first available spot on everyone’s calendars. It understands that it needs additional context to make sure a recurring weekly meeting will consistently work for everyone.

To do that, the agent might collect and analyze data and constraints like:

  • Past meeting patterns.
  • Individual time zones for remote teams.
  • Priority of the meeting relative to others on the calendar.
  • Alternative scheduling options.

The agent’s goal is to find the best options, so it will weigh these options and constraints to find the best choice.

Depending on how the agent is constructed, it may be running a planning algorithm to structure its tasks in a logical sequence. Reasoning models like Tree of Thought (ToT) or Reasoning + Acting (ReAct) are likely generating and evaluating options for the agent. The agent also uses Application Programming Interfaces (APIs) to gather data from various sources like calendars and CRM platforms.

3. Agents make decisions and respond to feedback.

After ingesting and analyzing data, the AI agent decides on an optimal date and time for the recurring weekly team meeting. So long as it’s running the right APIs, the agent can automatically build the meeting invite and send it to all parties.

The real agentic magic starts happening at this stage.

Let’s say the agent chose Wednesday at 4:00 PM for the recurring meeting. But, one team member, Alan, has to pick up his kid from daycare by 3:30 PM every day, and he didn’t add that to his calendar. So, he rejects the meeting invite.

Instead of ending operations, the AI agent learns based on feedback. When Alan says he couldn’t make this time, the agent automatically reassesses availability using this new constraint data. The agent selects a new meeting time and resends invitations to the marketing team. It picks Wednesdays at 1:00 PM, and Alan can make that work.

4. Agents execute tasks autonomously.

During this schedule preparation process, the AI agent is acting of its own accord. Think of all the tools or systems it might touch to handle this request:

  • Google Calendar or Outlook to check availability.
  • Slack or Email to communicate with the marketing team.
  • Zoom or Teams to set up a meeting room.
  • CRM tools like HubSpot to log team interactions.

The agent isn’t just offering a list of dates and times; it’s handling the entire scheduling process.

By calling functions and data through APIs, the agent interacts with other software to accomplish its objective without human intervention. Depending on the objective’s complexity, an agent might even take “initiative” and decide what external tools it needs to do the job and set up the integrations accordingly.

5. Agents remember and adjust based on context.

Now, it’d be easy enough to set it and forget it. The meeting is scheduled, the team is happy, and things are going great. However, an agentic AI can continue its work to help ensure long-term success with its tasks.

Not every AI agent has longer-term memory and context awareness. But of those that do, they can use that information over time to help your marketing team make better decisions.

For instance, this scheduling agent could remember Alan’s daycare needs and store historical meeting patterns as the weeks pass. It can apply that data to future scheduling needs.

In AI parlance, you’re no longer running a “stateless” operation, where AI handles only one prompt at a time. Instead, the agent stores pattern data in long-term memory frameworks like vector databases for later recall. Some agents even include episodic memory, which remembers past interactions for each user (e.g., Alan’s daycare needs).

6. Agents learn, adapt, and self-correct.

Over time, an AI agent refines its own processes to establish greater efficiency. For our scheduling AI, it would monitor the meeting and gather additional feedback to recommend adjustments.

It could track which times get the highest acceptance rates or how many times the meeting gets rescheduled and refine its logic over time. This mirrors Reinforcement Learning from Human Feedback (RLHF) but is closer to real-time optimization through adaptive learning models.

The AI then improves its ability to predict the best meeting times to reduce conflicts and optimize efficiency. It learns from its “mistakes” and self-corrects to do better next time.

7. Agents collaborate with other agents.

For our scheduling example, one AI agent is probably sufficient. But it’s possible for the scheduling agent to encounter other AI agents, such as one that handles email replies or manages project deadlines in your CRM.

A multi-agent system (MAS) requires collaboration between two or more agents to complete a common objective, much like a human team. These agents often chat with each other using structured coordination frameworks like decentralized reinforcement learning or hierarchical planning.

As AI gets more deeply integrated into companies’ workflows, I think we’ll see more opportunities for AI agents to delegate and negotiate tasks within a MAS.

When do I use an AI agent?

AI agents offer tremendous power and opportunities to any business. However, you also need to consider how you want to apply that power and what safeguards you install to monitor and adjust agentic AI’s use.

To explore this idea, Hilan Berger, COO of digital transformation consulting firm SmartenUp, shares his breakdown of agentic AI considerations.

“One of the first considerations is task complexity and scope. The complexity of the task determines whether a straightforward rules-based system will suffice or if a more advanced machine learning model is necessary,” he said.

“Another crucial factor is the autonomy level you require from the AI agent. Some AI solutions need to operate independently, while others serve as decision-support tools that work alongside human users. An AI’s adaptability and learning capabilities are also significant considerations,” Berger added.

“If the problem requires continuous learning and refinement, you’ll need a model with self-learning capabilities. On the other hand, a predefined rules-based system may be enough.”

Berger makes sure to highlight the human’s role in agentic AI. “You should always take into account decision transparency and compliance, particularly in regulated industries,” he said. “If AI-generated recommendations need to be auditable, like in financial forecasting, the system must provide explainable outputs.”

Pro tip: How else are marketing teams using AI right now? Check out our latest AI Trends for Marketers report for more details.

7 Types of AI Agents

While my scheduling agent example can show you the AI ropes, I should say that not all AI agents are created equal. In fact, most are built with intention and care to accomplish specific tasks and objectives.

We haven’t quite reached the stage where AI agents can truly act on their own (more on that later), but recent advances in agentic AI promise a fascinating future.

Let’s dive into the types of AI agents you might encounter now or later and how they can help your company.

Reactive Agents

If you watched an early model of a Roomba run itself into a wall, you’ve seen reactive agents in the real world.

Reactive agents are highly rules-based AI tools. They have a pre-programmed set of responses they adhere to rigidly, without the capability to learn from experience.

Reactive agents in business are excellent for automating low-level tasks that require basic repetition with predictable outcomes. You often see reactive agents operating as basic chatbots integrated into a website or in a workflow.

For instance, a sales-focused reactive agent would engage when a customer abandons their cart. The agent follows a conditional logic tree to “decide” what to do next, like sending a personalized email or text about the item left in the cart. AI-powered customer service and spam filters are also great examples of reactive agents.

Limited-Memory Agents

Limited-memory AI agents analyze recent data to make decisions, but they don’t store long-term knowledge (hence, “limited” memory).

This operational build works for tasks where you need up-to-date information but not long-term retention. For example, autonomous vehicles’ onboard AI makes real-time decisions based on current road conditions. That data should be consistently refreshed, so it’d be a waste of resources for the agent to hold onto it. You also see limited-memory agents in recommendation engines, like Spotify’s music recommendations.

Pro tip: HubSpot’s Breeze has AI that operates as a limited-memory agent, using your freshest HubSpot data to autonomously produce content, handle social media, conduct prospecting, and more. See what Breeze AI can do for your business.

Task-Specific Agents

True agentic AI operates with a lot of flexibility and decision-making capabilities. However, you sometimes have clearly definable high-volume tasks where AI could make a huge difference. This is a task-specific AI agent’s domain.

These agents are built with a highly narrowed and tightly defined purpose. For instance, Thomson Reuter’s CoCounsel AI serves as an AI-powered legal research agent to prepare legal work for clients. Coding assistants like GitHub Copilot or Amazon CodeWhisperer can suggest edits to code and run tests to validate updates.

Multi-Agent Systems

I touched on multi-agent systems earlier, but for more context, these systems involve multiple AI agents working together to accomplish a task. They truly lean into the concept that “the whole is greater than the sum of its parts.”

Industries like stock trading can benefit greatly from multi-agent systems. Multiple models could gather information from various sources, exchange data and insights, and collaborate to make more informed trades.

Multi-agent systems also have interesting physical applications. For example, a swarm of AI drones could deploy into a disaster zone and work together on search-and-rescue missions.

You’re unlikely to need multi-agent systems yet, unless you’re operating in specialized industries. But as agents proliferate, they’ll eventually come into contact with each other. It’s best to stay informed as agentic AI expands.

Autonomous AI Agents

It’s always a good idea to keep a human involved in any AI operation. However, when successes mount, you may start letting machines do more of the lifting. Enter the autonomous AI agent.

These agents operate with high autonomy, often optimizing processes or executing tasks on behalf of humans. Long-term memory and context help autonomous agents complete their objectives efficiently and adjust approaches based on past actions.

In the business world, you’ll see autonomous agents operating in departments like sales. Tools like Conversica automate significant chunks of the sales pipeline, and Salesforce’s Agentforce acts autonomously on various Salesforce-related tasks.

Theory of Mind Agents

Understanding data is one thing, but understanding human emotions is an entirely different realm. As advanced AI agents learn to work together, it’s possible they’ll learn how to understand the desires, behaviors, and attitudes of other agents — and humans — and predict how those mental states influence decisions and outcomes.

These “theory of mind” (ToM) agents cross the emotional divide between a machine and a person.

ToM agents are still in development, so don’t expect an immediate integration into your business. However, companies like Hume AI and Replika have built “affective AI chatbots,” which simulate human-like conversation, even if they don’t “understand” emotions yet. Woebot operates in the mental health space using AI therapists that can detect emotional patterns in a patient’s language and adjust responses accordingly.

replika theory of mind agent

Source

As the need for intelligent agents grows, ToM agents will serve as important partners for collaborating with (or competing against) other agents to accomplish more complex tasks.

For example, in the future, a ToM agent used by a consumer stock trading firm could infer a customer’s spending habits, risk tolerance, and motivations when monitoring trades. If a user is normally conservative but then suddenly makes several high-risk trades, the AI might be able to flag it as emotionally driven behavior and proactively suggest risk-mitigating actions like pausing trades or seeking a qualified financial advisor.

Self-Aware Agents

To be clear: Self-aware agents are still only hypothetical. While the U.S., China, and other countries are investing significantly in developing artificial general intelligence (AGI), self-awareness is not necessarily a requirement for AGI.

Perhaps the most famous fictional self-aware agent is Skynet — the killer AI that annihilates humanity in the Terminator franchise. It makes for classic cinema but doesn’t likely represent reality.

If self-aware AI were to emerge, it could function with a sense of its own existence, influencing how it makes decisions and interacts with us. Regardless of its intentions, the proliferation of self-aware AI would usher in another industrial revolution and upend how we think about work, society, and life itself.

How far away are self-aware agents? Benchmarking self-awareness is a science unto itself, and advanced AI agents are already sparking important ethical discussions on agentic AI’s applications. While I wouldn’t expect self-aware agents to join your office anytime soon, it’ll be an area to watch in the coming years (or decades).

Which AI agent is right for me?

Agentic AI is a developing field; what’s currently offered might not perfectly fit your needs. But, as you plug AI into your workflows, you’ll probably find a need to evolve your agentic AI choices over time.

“Businesses must assess whether they need a reactive AI that follows predefined rules, a limited memory AI that learns from past interactions, or a more advanced AI capable of adapting to new inputs in real-time,” said John Reinesch, Founder of digital marketing consulting firm John Reinesch Consulting.

“For example, in customer service, a company might start with a rule-based chatbot that answers common inquiries using predefined responses. This works well for simple, repetitive tasks but struggles with more complex or nuanced requests. As customer needs evolve, the business might shift to a machine learning-based AI that can analyze past interactions and adjust responses based on user behavior and sentiment,” he said.

I’d encourage you to have your team monitor AI use for opportunities and limitations within your current architecture. More advanced AI agents typically require more IT resources or larger AI experimentation budgets. Coming up with a solid implementation plan for agentic AI can help you convince leadership to increase investments.

Prepare for the Agentic AI Future

I’ve been cautious about AI’s integration into professional workflows. Yet the tools available today have impressed me with their capabilities. In practiced hands, you can accomplish a lot with AI.

If agentic AI fully comes to pass, I think it’ll feel like another quantum leap in reshaping work. While these tools evolve, the best way to prepare is to understand your company’s workflows and identify your team’s greatest needs. Prioritizing objectives and crafting a high-level implementation plan will get your team thinking ahead to integrate agentic AI effectively.

The future is agentic. Will you be ready?