Agents are back

Everybody is talking about “Agentic AI” all of a sudden. But the concept of software agents has been around for a long time, since the actor model of software computation [1] from the 1970s. So why, after all this time, are they making a comeback? Rafa Bren shed some light on this here. 

Agentic AI refers to Artificial Intelligence (AI) systems designed to operate autonomously, make decisions, and take actions on behalf of a user through the use of software agents. An agent is typically defined as an entity that can take actions to achieve specific objectives. What sets AI Agents apart from traditional AI systems is their ability to act with a level of autonomy, actively pursuing goals in dynamic, real-time environments rather than responding to input or following programmed instructions.

The principal drivers of this resurgence of interest are advances in AI brought about by large language models (LLMs). Armed with much improved abilities to generate and understand instructions in natural language, software agents can reliably communicate and collaborate on tasks through use of services with clear application programming interfaces (APIs), and increasingly, user interfaces. This is particularly relevant in the case of multi-agent systems, where multiple agents have to collaborate in order to accomplish a task. Indeed, a number of frameworks for orchestrating AI Agents, such as Crew AI, Semantic Kernel, AutoGen, LangChain, and LlamaIndex have emerged.   

In AI Agents in Action, Michael Lanham characterises 4 interaction types with a LLM:

1. 1. Direct interaction, e.g. “Please explain the definition of an agent”. 
      
      
   2. Agent/assistant proxy: reformulation of user requests to other services, e.g. “Show an illustration of an agent”. 
      
      
   3. Agent/assistant: acting on behalf of the user with approval, e.g. “What is the temperature today in Calgary today?”. 
      
      
   4. Autonomous agent: making decisions on behalf of the user, e.g. “Filter my emails by importance and notify me of the top 5 most important emails”. 
      
      

Many people are familiar with these types of example interactions with computer assistants before the world of LLMs. The brave new world of LLM-powered agents presents the opportunity to decompose much more complex problems into smaller tasks that can be orchestrated. Further technical innovations with LLMs have helped to enable this revival:

1. 1. LLM as a knowledge base: customising a foundational LLM model by fine-tuning it (changing the parameters of the model) with domain-specific data. 
      
      
   2. Retrieval augmented generation (RAG): augmenting responses from a foundational LLM with context from domain-specific content.

Customising LLMs to encapsulate domain knowledge to address vertical specific problems is powerful, especially when they can be composed. While both Fine-tuning and RAG attempt to solve the same problem, RAG has emerged as the more viable approach for many use cases [2].

Two other technical concepts are worth noting here:

• Neuro-symbolic AI is the combination of LLMs with symbolic reasoning techniques used in more traditional AI systems.This enables the LLM to defer certain tasks to more traditional software components [3]. 
  
  
• A Large action model (LAM) is an LLM specifically trained to execute actions from data, often utilizing a neuro-symbolic approach [4]. 
  
  

Automation opportunities

Agentic AI presents a huge potential opportunity in the technology-enabled automation of business processes due to this resurgence. Robotic Process Automation (RPA) companies have been quick to adopt the technology. New business value will be created in addition to operational efficiencies.  

At a glance, a high level characterisation of solutions is as follows:

1. 1. Augment human process with digital tools (“augmented worker”). 
      
      
   2. Encapsulate domain knowledge and subject matter expertise in software services (“expert in a box”). 
      
      
   3. Create digital workers with human oversight (“human in the loop”).

The future of work

A key question to answer in the coming years is what the impact of these technological advancements will have on jobs?

A recent report from Mckinsey predicts that by 2030, up to 30 percent of current hours worked could be automated with the advent of AI, as well 12 million occupational transitions each in Europe and the United States. They also predict that businesses will need a major skills upgrade and occupations with lower wages will likely see a reduction in demand [5]. 

But the global economy has seen seismic shifts in technology before. The World Economic Forum notes that while new technology often replaces certain jobs, it also creates new roles. Farm technology saw people move from agriculture, the industrial revolution moved people to factories, while automation gave rise to the service economy. Today there is a record number of people in employment across the globe [6].

Towards fully autonomous systems?

As LLMs become more and more sophisticated, they can still suffer from quality issues at times. For example, texts generated via LLMs often suffer from accuracy issues including omission of key details, confusion of inputs (e.g. wrong attribution of information to entities), and hallucinations (pure factual errors).

Software agents also run the risk of manipulation, amplifying errors / bias and clouding accountability for decisions. For this reason, guardrails will continue to be applied and human oversight should remain in place for many implementations. This should not detract from the overall opportunity however, as the benefit should far outweigh the cost of any necessary safety measures that need to be taken.

 

3 things to take away:

1. 1. Agentic AI is not a new concept, but advances in large language models means that machine to machine communication via natural language opens up new possibilities for automation. 
      
      
   2. The automation of work outlook is still bright, while jobs may be automated, new ones will be created. 
      
      
   3. The promise of full automation is unlikely for now. Trust and quality requirements will persist the need for guardrails implemented via human oversight.

1.  

 

[1] Actor Model

[2] LLMs as a knowledge base vs LLMs with Knowledge Retrieval

[3] Jurassic X Crossing the Neuro-symbolic Chasm with the MRKL System

[4] Large Action Models

[5] A new future of work: The race to deploy AI and raise skills in Europe and beyond

[6] Why there will be plenty of jobs in the future — even with artificial intelligence

[7] Image attribution: “A futuristic Agentic AI secret agent team: Multiple sleek humanoid robots wearing high-tech trench coats and futuristic sunglasses, each equipped with advanced AI gadgets like holographic briefcases, wrist-mounted data interfaces, and stealth drones. They are gathered in a dimly lit cyberpunk city alley, strategizing a covert mission. Some agents are blending into the shadows while others scan their surroundings with glowing blue eyes. The scene conveys teamwork, intelligence, stealth, and advanced decision-making capabilities, evoking a high-tech spy thriller atmosphere.” prompt, DALL-E, version 3, OpenAI, 26 Mar. 2025, https://chatgpt.com/