AI Gets Promoted: From Assistant to Autonomous
Understanding the upcoming generation of AI agents that can think, plan, and act independently, and why it matters for teams and businesses
It’s been an incredibly busy few weeks delivering client projects and workshops. And while I have thoroughly enjoyed myself, it has also meant time away from one of my favourite activities – writing about AI. So I’m glad to have found a few moments this week to jot down my thoughts about AI Agents, a topic of growing interest.
On a separate note, my AI Foundations for Freelancers course will be running over three punchy lunchtime sessions on 27 November, 05 December, and 11 December. Our early bird offer (25% off!) is available up to 14 November. If you’re a freelancer, small business owner, or a business professional looking to power up your AI game, I invite you to reach out to me at justin.tan@evolutioconsultingco.com to learn more.
"Use data from my computer and online to fill out this form."
Until recently, this simple instruction would have required human intervention at multiple steps – checking spreadsheets, opening browsers, navigating websites, and manually entering data. It was like asking someone to make you a sandwich but having to tell them exactly how to open the bread bag, how many steps to take to the refrigerator, and which way to spread the mayonnaise.
But with Anthropic's recent introduction of "Computer Use" capabilities for their AI assistant Claude, we're seeing the emergence of AI Agents that can perform these tasks autonomously, just as a human would. Indeed, Claude could now use computers just as humans do – moving cursors, clicking buttons, and navigating interfaces.
While this might seem like a modest advancement, it represents something far more significant: the emergence of autonomous AIs that can interact with and manipulate their environment to achieve specific goals. For the parents out there, the analogy I can think of is watching your teenager finally figure out how to do their laundry without turning everything pink (true story!) – a small step that signals a much bigger leap toward independence.
While traditional AI models require specific instructions, AI Agents assess situations, plan sequences of actions, and work toward defined goals with minimal supervision. Think of a kitchen assistant (e.g., me!) moving up the experience curve to become an executive chef (e.g., my partner!) – no longer just following recipes step by step, but understanding the principles of cuisine, adapting to available ingredients, managing kitchen dynamics, and even improvising new dishes when needed.
This shift has massive implications for businesses across every sector. As we stand at the cusp of this new era, business leaders need to understand what AI Agents are, how they differ from traditional AI systems, and the transformative impact they could have on their organisations. This article will explore the fundamentals of AI Agents, their various types, and their practical applications in business contexts.
The Evolution from Assistant to Agent
Think of how we develop talent in organizations – from closely supervised newcomers to trusted autonomous decision-makers. AI has followed a similar path.
It started with Robotic Process Automation (RPA) in the 2000s – digital workers that excel at repetitive tasks but follow rigid rules. Like that one colleague who's brilliant at their specific job but gets flustered when the printer jams, RPAs are unable to work with anything that is at all unexpected.
Next came AI assistants – from basic ones like Siri and Alexa to more capable platforms like ChatGPT and Claude in the last two years. They're rather like eager interns: smart and capable of drafting emails, analysing data, or finding information, but popping by your desk every ten minutes to ask, "Does this look right?" While impressive, they remain fundamentally reactive.
Today's AI Agents represent a significant leap forward. Unlike the rigidity inherent in RPAs or the reactive nature of AI Assistants, these agents are capable of managing entire workflows independently.
What makes these agents truly capable? They can:
Grasp and interpret high-level objectives
Break complex tasks into manageable steps
Navigate different tools and interfaces
Learn from their interactions
Adapt decisions based on changing circumstances
This evolution stems from breakthroughs in several key areas:
Advanced language understanding: Modern Large Language Models grasp nuanced instructions and context
Environmental perception: New capabilities allow agents to "see" and understand digital interfaces
Action planning: Sophisticated algorithms enable multi-step task planning
Tool manipulation: Direct interaction with software through APIs and other approaches
Learning systems: Decision-making refined through iterative feedback
Assuming AI Agents deliver upon their promise, humans will eventually be freed up to focus on what they (should) do best – strategic thinking, creative problem-solving, and working on ‘human-to-human’ challenges.
Types of AI Agents
If AI agents were a corporate workforce, you'd find them filling roles from junior associates to C-suite executives. Each type has its own specialty, much like how different roles in an organisation contribute to the bigger picture. Let's explore this hierarchy of artificial talent, starting from the ground up.
Simple Reflex Agents: The Dependable Specialists
Picture a highly focused junior analyst who excels at specific tasks but doesn't need to understand the broader context. These agents are the specialists of the AI world – they see a specific situation and know exactly what to do, no questions asked. These agents are similar to traditional RPA but with more sophisticated “condition-action” triggering mechanisms.
In business settings, they excel at routine but essential tasks: monitoring system alerts, flagging unusual transactions, or triggering automated responses to specific events. For example, in cybersecurity, these agents continuously monitor network traffic and respond instantly to known threat patterns. They're very efficient at what they do, but don't ask them to improvise – that's not their strong suit.
Model-Based Reflex Agents: The Experienced Operators
These are your seasoned middle managers who've been around long enough to understand how their departments work. They don't just react to situations – they maintain a mental model of their environment, drawing on both current information and historical context to make decisions.
Take a smart building system for instance. It doesn't just react to current temperature readings; it considers time of day, typical occupancy patterns, weather forecasts, and energy costs to optimize comfort and efficiency. In customer service, these agents can handle complex interactions because they understand customer history, common problem patterns, and typical resolution paths – much like an experienced service representative who knows when to escalate issues and when to handle them directly.
Goal-Based Agents: The Problem Solvers
When it comes to Goal-Based Agents, think project managers and tactical planners. While Simple Reflex and Model-Based Reflex Agents operate on "if-this-then-that" logic, Goal-Based Agents ask "What do I need to do to get from A to B?”. They evaluate multiple pathways to achieve specific objectives, weighing and iterating through options as a chess player would.
In financial services for example, Goal-Based Agents might manage investment portfolios. Rather than simply following market triggers, they actively plan trading strategies to meet return targets while navigating risk parameters. They might choose between aggressive growth tactics in bull markets, defensive positions during downturns, or take opportunistic bets, while always keeping the end goal in sight.
Utility-Based Agents: The Executive Decision Makers
If Goal-Based Agents are project managers, Utility-Based Agents we can likened to seasoned executives who excel at balancing complex trade-offs. They don't just pursue single objectives but weigh multiple competing factors to find optimal solutions.
In hospital networks for example, it's not sufficient to simply reduce wait times (a single goal). Agents working in such an environment would need to balance emergency response capabilities, specialist availability, equipment utilisation, and patient priorities across multiple facilities. Every decision involves trade-offs between individual patient outcomes and system-wide efficiency, while rapidly adapting to changing medical priorities.
Learning Agents: The Adaptable Innovators
These are your organization's rising stars – the ones who not only perform their tasks but continuously improve through experience. Like promising employees who grow beyond their initial training, Learning Agents refine their decision-making and discover better ways to achieve their objectives.
In manufacturing for instance, such agents could have the ability to optimise production processes by analysing performance data and adjusting parameters to improve efficiency. What makes them valuable is their ability to adapt to changing conditions without requiring reprogramming.
Multi-Agent Systems (MAS): The High-Performing Teams
This is where things get really interesting. While individual agents are powerful, the real magic happens when they work together, each bringing their unique capabilities to solve complex problems. It's like watching a well-oiled corporate team where specialists collaborate seamlessly toward common goals.
Without explicit programming for social behavior, these agents developed uncannily human-like interactions. They noticed each other, formed relationships, organised gatherings, and even coordinated party planning. It's a fascinating glimpse into how multiple AI agents can create emergent behaviors that mirror human social dynamics.
MAS can be structured in several ways:
Hierarchical systems where "manager" agents coordinate "specialist" agents
Peer networks where agents collaborate directly, like cross-functional teams
Competitive systems where agents pursue individual goals while maintaining system balance
Hybrid approaches combining these different organisational structures
The real breakthrough comes when different types of agents work together in modern business systems – Simple Reflex Agents handling immediate responses, Model-Based Reflex Agents providing context, Goal-Based and Utility-Based Agents making tactical and strategic decisions, and Learning Agents continuously improving system performance. Wait. That sounds like an organisation that actually works. Gasp.
For those interested in getting under the hood of such systems, I’d encourage you to look into one of my favourite MAS experiments. Scientists create a virtual neighbourhood of 25 AI Agents, sort of like “The Sims” but with a twist.
AI Agents Today
Most AI Agents today fall into two main categories: Simple Reflex Agents and Model-Based Reflex Agents.
Simple Reflex Agents are commonly seen in basic automation tools, such as chatbots, rule-based customer service systems, simple trading, or spam filtering systems. Model-Based Reflex Agents, on the other hand, can be found in more advanced technologies like autonomous vehicles, smart home systems, and industrial control systems.
Here are some examples of such AI Agents in play today:
The Humble Hero: Remember Roomba, that disc-shaped vacuum that bumps around your home? It's actually a Simple Reflex Agent in disguise. When it hits a wall, it changes direction; when it detects dirt, it starts cleaning. Newer models have evolved into Model-Based Reflex Agents, mapping your home and remembering cleaned areas to optimise their routes.
The Silent Guardian: Siemens' Cyber Defense Center (CDC) tackles the herculean task of processing 60,000 potentially critical security events every second, and is an example of a Model-Based Reflex Agent. It autonomously labels data, trains machine-learning models, and responds to real-time threats. The system is so effective it needs just a small team of humans to oversee operations.
The Road Warriors: Waymo's self-driving vehicles showcase Model-Based Reflex Agents at their finest. "Eyes and ears" come in the form of multiple sensors including external audio receivers, cameras, LiDAR, and radar. The onboard AI processes this data and acts fully independently through the lens of experience gained from 20 million real-world miles and 20 billion simulated miles.
Glimpses of the Future
While most current AI agents are relatively simple, more sophisticated versions are beginning to emerge. Here are some examples:
The Game Changer: AlphaGo, developed by DeepMind, is one of the earliest examples of Learning Agents. By playing millions of games against itself, it learnt and developed strategies that helped it defeat world champion Lee Sedol in 2016. My article here goes into more detail about AlphaGo and how it learns over time.
The Digital Developer: Cognition AI's Devin AI, developed by Cognition AI, is a software development agent that can write, test, and implement code solutions. In tests by Bloomberg, it built a website in ten minutes and tackled computer vision projects from Upwork. A later version introduced multi-agent capabilities, with ten 'worker' agents coordinated by a manager agent, akin to a small development team led by a technical lead.
The Strategic Commander: In the complex world of StarCraft II (a real time strategy game that makes chess and Go look simple and of which I’m unashamedly a big fan of!), AlphaStar learned to manage hundreds of units simultaneously while processing incomplete information, very much like a military commander making rapid decisions in fog of war conditions.
These more advanced agents, while impressive, still face significant challenges. Devin AI's 13.9% success rate on basic coding tasks and Anthropic's Computer Use achieving 14.9% on basic computer tasks (compared to human scores of 70-75%) remind us that we're still in early days.
That being said, AI Agents have come very far in a short period of time. Given the billions of dollars going into this space and thousands of talented teams working on them, I believe its only a matter of time, perhaps as close as 12-18 months before we begin to see the first Goal-Based and Utility-Based Agents deployed at scale in early adopter enterprises.
The Pipeline
While the list of sophisticated and fully autonomous AI Agents in real world use is currently limited, tech companies aren’t standing still. Here’s what’s cooking:
The Universal Butler: Google's Project Astra (see video demo below), announced earlier this year and targeting a 2025 release, aims to be a universal AI Assistant with real-time visual processing, contextual memory, and advanced planning capabilities. Think of it as a very advanced version of Siri or Alexa that can see, remember, and plan complex tasks while handling interruptions naturally.
The Nano Navigator: Aitomatic's Open Small Specialist Agents (OpenSSA) architecture targets the semiconductor industry with specialised decision-making capabilities. Built on Large Language Models (LLMs) that are specific to the semiconductor industry, their agents will be able to optimise and manage complex manufacturing processes (though I'll spare you the technical details about plasma etching parameters!).
The Grid Maestro: AES, an American utilities company, and Google's Tapestry project are developing AI agents for electric grid management, moving beyond current machine learning applications toward autonomous orchestration of grid operations and management of real-time market transactions.
Business Applications and Impact
While the tech world buzzes about AI Agents, the real question is: what can they do for your business in the near future? Let's explore how these digital workers could transform different sectors, focusing on practical applications within the next 12-18 months.
Healthcare
Healthcare systems worldwide are groaning under increasing pressure. AI Agents offer a way to expand capacity and improve care quality without burning out staff:
Patient Support Agents: Automate routine patient inquiries and allowing staff to focus on more complex cases. This is what Germany-based Avi Medical has used Beam AI's multilingual AI agents to achieve, helping reduce response times by 87% while boosting patient satisfaction by 9%.
Drug Discovery Agents: AI models like Alphafold are already reshaping how we discover new drugs. Add autonomous agents to the mix, and researchers will be able to significantly shrink the typical 10-15 year development cycle by having them identify promising compounds, predict efficacy rates, and simulate potential outcomes around the clock.
Virtual Health Agents: Imagine having a dedicated coordinator for each patient, independently managing appointments, providing personalised medication reminders, and running pre-consultation information gathering. Such agents could help reduce those frustrating wait times while ensuring doctors have better patient information before consultations.
Administrative Agents: Here's a sobering statistic – the average UK clinician spends one-third of their time on documentation. AI Agents could reclaim these hours for patient care by streamlining administrative tasks, such as managing records, handling billing, and automating data entry tasks.
Financial Services
The Financial sector has long been a heavy user of AI models for fraud detection and prevention, investment advisory, customer services, and risk assessment and credit scoring. Here’s what’s next:
Financial Planning Agents: Unlike today's static robo-advisors (e.g., Wealthify, Nutmeg) that simply rebalance portfolios based on preset rules, the coming wave of agents will be more like personal financial planners and coaches. They would adapt and execute plans based on real-time life events such as job changes, major expenses, market shifts, making personalised financial guidance available to all.
Negotiation Agents: Ever thought of having a skilled negotiator working 24/7 to secure the best terms for your loans, insurance, or mortgages? These agents could bring efficiency, speed, and transparency to traditionally opaque markets, potentially disrupting the current intermediary-heavy model.
Trading Agents: While the algorithmic trading bots of today tend to follow rigid rules ("if X happens, do Y"), more sophisticated trading agents could adapt formulate and adapt strategies based on complex market conditions, geopolitical events, and economic indicators, more akin to seasoned traders than simple automation tools.
Retail / Ecommerce
AI models are currently employed by Retail and Ecommerce sector companies for personalising product recommendations, automating basic customer service interactions, and for inventory and supply chain management and dynamic pricing. Autonomous agents could potentially be employed in the following ways:
In-Store Experience Agents: These wouldn't just be glorified store directories. Think personal shopping butlers that understand your style, budget, and preferences, guiding you through stores while answering detailed product questions. The human touch will likely always be welcome in retail, especially for building trust and customer relationships, and such agents would allow human staff to focus on handling complex interactions and those requiring empathy.
Logistics Agents: We would move beyond basic route optimisation to agents that orchestrate entire delivery networks. Such agents would coordinate multiple transport modes, negotiate with logistics providers, and proactively communicate and discuss with customers about delivery options – all sans or with only minimal human intervention.
Inventory Management Agents: Beyond simply forecasting and managing inventory, these agents could autonomously coordinate across multiple stakeholders, including suppliers and logistics partners, anticipate supply chain disruptions and proactively secure alternative sources. They could also integrate real-time data from customer feedback, sales forecasts, and even market conditions, ensuring a highly adaptive and resilient inventory strategy.
The Next Chapter
The story of AI Agents will unfold differently across sectors, geographies, and organisations. But one constant remains: we're entering yet another significant shift in how we work.
Just as we learned to collaborate with ChatGPT and Claude as 'copilots' – mastering prompts, understanding quirks, accepting limitations – working with AI agents will demand an even steeper learning curve. We'll need to:
Define boundaries while allowing agents room to operate
Monitor outcomes rather than micromanage processes
Develop "agent oversight" skills – knowing when to intervene versus letting agents learn from mistakes
Build comfort with digital delegation, especially where perfection isn't critical
Create feedback loops that help agents improve over time
And it's not too early to start. While fully autonomous agents are still evolving, the tools for building them have already come a long way. Developers can explore platforms like Autogen and Crew AI, while those preferring a code-free approach can use Microsoft's Copilot Studio to create simple agents. These platforms offer low-risk entry points to understand how agents think and work.
The smart play? Start small in lower-risk areas. Begin with agents handling internal processes where mistakes are manageable, or in contained environments where you can easily monitor and correct outcomes. Use these experiments to build organisational muscle memory.
Remember those clunky first conversations with ChatGPT, and how you gradually learned to extract better responses? Working with AI agents will follow a similar pattern – but at an organisational scale. The winners won't be those with the most advanced agents, but those who most effectively learn to work alongside them.
The next 12-18 months aren't about perfect deployment – they're about perfect learning. Every stumble and success will shape how your organisation integrates these new digital colleagues. Don't worry about getting it right – focus on getting started!
Justin Tan is passionate about supporting organisations and teams to navigate disruptive change and towards sustainable and robust growth. He founded Evolutio Consulting in 2021 to help senior leaders to upskill and accelerate adoption of AI within their organisation through AI literacy and proficiency training, and also works with his clients to design and build bespoke AI solutions that drive growth and productivity for their businesses. If you're pondering how to harness these technologies in your business, or simply fancy a chat about the latest developments in AI, why not reach out?