What are AI Agents?

Today, AI agents are driving the biggest shift in technology since the invention of the smartphone, changing how we interact with computers in a huge way. Let’s take a quick look at what AI agents are, where AI agents stand today, the challenges we face making them useful, how they change the workplace and where they are heading next.

Definition

An AI agent is an autonomous software system designed to achieve a specific goal by reasoning, planning, and executing actions on its own. Yeah, but what is it really?

Unlike traditional AI chatbots that simply answer a single user prompt and wait for the next command, an AI agent operates in a continuous loop, the benefit being it is always on if you need it to be. The agent evaluates a task, breaks it down into a plan, executes actions using external software tools, observes the results, and adjusts its approach until the job is completed. Let’s say it hones in on the problem to come up with a solution. Hence the ‘autonomous’ part of the definition.

The 4 Core Components of an Agent

To operate autonomously, AI agents rely on four foundational layers:

1. The Brain (LLM): Large Language Models (like OpenAI’s GPT, Google DeepMind’s Gemini, or Anthropic’s Claude) serve as the core engine. The model handles the central reasoning, logic, and intent classification.
2. Planning & Orchestration: The agent splits complex goals into smaller, sequential steps. If a strategy fails, it can self-correct, dynamically rewriting its to-do list based on real-time feedback.
3. Memory: Agents use short-term context windows to track immediate work logs, combined with vector databases as long-term memory to recall past actions, user choices, or historical company records.
4. Tools (The Hands): This layer connects the AI to the physical or digital world. Agents are given access to custom APIs, web browsers, calculation scripts, email clients, or terminal environments to fetch data or complete transactions.

Real-World Examples

• Customer Service: Managing end-to-end user issues — such as looking up an invoice, validating a return policy, processing a refund, and emailing a shipping label — completely without human intervention.
• Enterprise Operations: An automated system routing specialized inquiries. For example, a legal agent might pull incoming claims, run a fast classifier to flag easy issues, and leverage a heavier research engine to synthesize case law for complex contracts.
• Software Engineering: Autonomous tools that don’t just suggest code syntax, but actively read an entire local repository, spin up environments, run automated tests, debug runtime errors, and submit pull requests.

Agent construction generally falls into three architectural (IMHO) patterns:

• Single Agent: One AI model handles the planning, tool use, and execution. This works well for straightforward, narrow tasks like organizing an inbox or scraping a specific website.
• Multi-Agent Systems: Multiple specialized AI agents work together like an enterprise team. For example, a software project might have a “Product Manager Agent” write the requirements, a “Coder Agent” write the software, and a “QA Agent” test the code for bugs and send it back if it fails.
• Human-in-the-Loop (HITL): The agent does 90% of the heavy lifting but automatically pauses and asks a human for approval before taking high-risk actions, like spending money, sending a bulk email to clients, or editing or deleting files.

Problems

While I love this poweful concept, making and engineering agents that really work is incredibly difficult due to a few major challenges:

“Infinite Loop” Problem

If an agent encounters an unexpected error (like a broken or reformatted website layout or an API timeout), the agent can get confused, try the same broken fix repeatedly, and run up massive cloud computing bills without finishing the task.

Cascading Errors Problem

If an agent makes a small logical mistake or hallucinates in step 2 of a 20-step plan, every subsequent step builds on that falsehood, causing the entire task to completely go off-track and by the end of the run, the task is in chaos or completely wrong and again waste compute cycles.

Security & Trust

Giving an autonomous system access to a real company database or credit card comes with major risks. Governance, compliance and guardrails are key. Developers have to strictly limit what “tools” the agent is allowed to touch and at the same time maintain observability into what the agent does touch.

Heavy Testing

A lot of effort needs to be put in upfront to ensure the agent does what it is supposed to. This is new technology and there aren’t a lot of experts available (yet) to craft these agents efficiently.

What’s Coming Next?

We are moving away from “text-in, text-out” AI and moving toward “intent-in, outcome-out” computing. In the near future, instead of using dozens of individual apps to book a vacation — checking flights on one site, hotels on another, and calendar availability on a third — you will simply tell a personal AI agent your budget and preferences. The agent will orchestrate the entire process across the web in the background while you do something else.

How will they change jobs?

AI agents are fundamentally changing the nature of work. The consensus among workforce experts (and realise the first LLM was released in 2018 so these experts are only a few years old) is that AI agents are not replacing entire jobs; they are replacing task layers.

A traditional job is made up of dozens of individual tasks. Because AI agents can operate across multiple software tools autonomously, they are starting to absorb the routine, data-heavy and administrative “middle steps” of a workday, such as
data entry, scheduling, summarizing, and basic code writing. I spoke to an attourney and her firm is using AI to perform legal research that interns used to do. She said the time savings was phenomenal but the accuracy still needs work so the interns still have to check the results.

The Core Skill Shift: From Writing to Reviewing

For white-collar workers, like the author, daily work is shifting away from building things from scratch and moving toward having a head-start and then auditing, framing, and directing.

Software Engineers

Instead of writing lines of code by hand, developers are becoming code reviewers. An agent drafts the code and runs tests, while the human engineer reviews the architecture, ensures security, and approves the final deployment.

Financial & Business Analysts:

Instead of spending hours pulling spreadsheets and matching data, analysts act as “value architects”. They set the parameters for AI agents to query corporate data and focus their time on strategic decision-making and investigating anomalies.

The Rise of “Agent Orchestration”

A brand-new category of high-paying jobs is emerging around managing these digital workforces. Enterprise hiring data reveals that job postings referencing “agent orchestration” exploded by 1,721% over the last year.

Companies are aggressively hiring AI Agent Engineers and forward-deployed engineers who are versant in deploying agents.

These professionals do not necessarily build the AI core from scratch; instead, they act like digital foremen — designing the workflows, connecting agents to corporate tools, setting up safety filters, and managing the cloud computing costs of the digital workforce.

The “Missing Bottom Rung” Dilemma

One of the most complex societal impacts of AI agents is happening at the entry level.

• Historically, junior employees learned their trades by doing the “grunt work” — summarizing documents, fetching data, or formatting initial drafts.
• Because AI agents can do these tasks instantly, data shows that hiring for junior and entry-level positions has become incredibly competitive and challenging.
• This leaves a critical gap: if agents do all the beginner-level tasks, how do companies train the next generation of mid-level experts?

What Makes a Job “AI-Proof”?

According to recent research and economic reports, including McKinsey Global Institute data, the jobs most resilient to AI automation rely on three major pillars:

The “Answering For” Rule

The dividing line is no longer how smart a task is, but who carries the legal, financial, or ethical liability. A machine can suggest a diagnosis or write a contract, but a human doctor or lawyer must sign off and take accountability for the final outcome.

Think about this: if a driverless car has an accident, who is accountable?

High Human-Touch Roles

Jobs dependent on deep empathy, trust, negotiation, and relationship-building — such as physical therapists, educators, or specialized counselors — remain highly human-centric and are at low risk of being ‘replaced by AI’ anytime soon.

Physical Adaptability

Roles requiring complex movement in a constantly changing physical space, such as electricians, plumbers, or field construction workers, are incredibly difficult to automate because robotics cannot yet match human physical problem-solving in unstructured environments.

Finally

While this new technology is thrilling and fascinating, we need to proceed cautiously and carefully as well as be empathetic to those who may be replaced. A friend of mine said there should be an AI tax to help those who are replaced by AI and I said, “I thought you were against taxes?” 😄

If you are interested in seeing what the code for an agent looks like you can go (here)[https://awindest.github.io/ai-agent-coding-project/].