Everything You Need To Know About AI Agents For L&D (2026)

Agent this, agent that is literally all I see on LinkedIn these days.

Granted, it’s a total echo chamber of people mostly shouting that back at each other, but by God, it’s giving me a headache.

The hype, mostly driven by AI companies, is becoming laughable.

Don’t get me wrong, AI agents will be very useful and there’ll be some great applications in L&D, yet you’d think a sort of world peace is about to emerge by the way the ‘influencers’ talk on social.

So, this is my PSA (public service announcement) to you to say: “Don’t get worried about all the talk.”

I know it feels like you’re missing out on some great party, but you’re really not.

Don’t believe the AI Agent hype in L&D

Almost 90% of what you see paraded online is not a true agent solution. Not in the technical context, anyway.

Much like marketing teams decided to use the word “AI” everywhere post-2022, they’re doing the same thing by labelling everything an Agent.

Unfortunately, this has created a fractured understanding of what an agent is, and the definitions are always changing.

Not only this, but many are trying to run before they can walk.

I work with sooo many teams and companies that hardly know how to use a basic AI assistant to even 50% of its potential. Adding agents into that mix is a recipe for both confusion and mistakes.

A lot of people need to slow down…

Pause… take a breath and find your centre (or whatever meditation teachers say).

Without this moment of pause, it’s incredibly hard for you to truly know what’s going to help you and what you don’t need to know.

And too many of us aren’t aware of all the options we already have today.

Agents are cool, but the current noise is lying to you about a lot of things.

So, let’s bring some clarity to all of this ↓

Assistants vs Agents: What’s the difference?

Two terms you might hear techies mention with AI products are ‘AI assistants’ and ‘AI agents’.

Here’s the difference in clear, simple terms.

Let’s start with what we know – AI assistants like ChatGPT.

These are tools that help us with tasks through conversation. They can write, analyse, explain, and give suggestions based on what we ask.

AI agents take this a step further.

Instead of just helping through conversation, agents can actually complete tasks on their own. They follow instructions, use different tools, and make basic decisions to get things done.

The key difference is simple:

• AI assistants help you with tasks
• AI agents complete tasks for you

Both are valuable, but they serve different purposes. An assistant works with you through conversation, while an agent works independently based on your instructions.

Use this info to impress the boss at your next meeting.

I’m not going to leave you with just this, though.

As I’m a tech nerd, I’ve filmed a quick video (see below) to show how agents work with examples from Google and Salesforce – enjoy.

What can AI agents do?

A lot, but maybe not as much as the local tech bros are promising.

Imagine having a personal assistant who not only follows your instructions but also takes the initiative to resolve problems independently.

AI agents are like that, except they exist in the digital world.

At their core, they’re designed to observe their environment, make decisions, and take actions using the tools available to them.

Unlike traditional software that waits for you to give it a command, like LLMs, AI agents can think ahead, figure out what needs to be done, and act.

Sometimes without needing constant human input.

Think of them as a self-driving car.

Instead of waiting for a person to steer, brake, or accelerate, the car analyses traffic, makes decisions, and moves safely toward its destination.

AI agents work similarly but in a digital space, whether it’s automating workflows, analysing data, or even assisting with creative tasks.

The magic of AI agents lies in their autonomy and problem-solving abilities.

Even if you don’t give them step by step instructions, they can work out the best way forward to achieve a set goal.

They do this by following set rules and past experiences to decide the best way to complete a task. This makes them incredibly useful for businesses, customer support, research, and even personal productivity.

→ Get an example of this type of AI agent solution with this scenario I built to support common onboarding challenges between HR and Tech teams.

The many faces of AI agents

There was a time when an AI agent meant one thing.

Now, we’ve hit peak confusion thanks to marketing teams the world over.

Each one wants to tell you they’re “agentic”, and each wants you to use their AI agent. But…is it really an AI agent? And if it is, is it the right one for you?

Let’s unpack the types of AI agents, or what social media wants to tell you are AI agents in the market today:

Now, the reality of what you see online is 95% in the automation and AI workflow buckets.

I know every 22-year-old with a YouTube channel wants to tell you otherwise, but “true” AI agent solutions, right now, are rare. Even rarer are agents doing valuable work within organisations.

And when I say ‘agents’, I mean actual ‘agents’, not workflows.

I’m not being harsh. I think AI workflows and automations are very useful, just don’t call them “Agents”.

Before we move on, let’s talk about Model Context Protocol aka MCP, in the first image.

Unless you’re a backend developer or some super nerd (like yours truly), you might never engage with MCP. Nonetheless, let’s take this as a learning moment to once again impress at your next team meeting.

Model Context Protocol Explained

To understand MCP, we need to understand the limitations of Large Language Models (LLMs) on their own, with the challenges developers face when trying to make them useful.

Maybe this will make you feel a bit of empathy for your local tech team.

LLMs are good at tasks like writing text, answering questions based on their training data, or generating code snippets.

However, they can’t do anything meaningful in the real world on their own, such as sending an email, interacting with a calendar, or performing a specific task on your behalf.

So, we need to connect them to different tools and services.

We can do this through APIs…however, this relies on APIs being made available for applications to connect and constantly needing to be monitored. One API with an LLM is easy, but connecting multiple tools to LLMs through APIs is difficult.

Now, MCP helps solve this problem by acting as a universal translator to simplify these connections.

Think of it as a layer between the LLM and all the different tools and services it might need to interact with. Something like a universal adapter. Instead of the LLM having to learn and manage every single service (through an API), MCP translates the different “languages” of all those services into a unified language for the LLM.

Now, either you got that, or I confused the s**t out of you.

If the latter, check out this vid, which should resolve that.

You probably already have (and used) an agent

You don’t need to go shopping for a standalone AI agent. The tools you’re already using have fast become agentic without screming “agent” in your face.

The line between “AI assistant” and “AI agent” I drew earlier is blurring fast, because the big AI companies have baked agentic capabilities directly into their everyday products.

Here’s some examples:

Claude

Anthropic now ships three flavours of agentic Claude:

• Claude Cowork: This lives as a tab in the Claude desktop app. You point it at a folder on your computer, describe the outcome you want, and it works through the task on its own. That means it can read files, create documents and organise the chaos in your downloads folder. No setup beyond “here’s my folder, off you go” and some well crafted instructions. You can also give Cowork access to use your browser on its own, yet I’d caution you on that one unless you’re cool with potentially exposing your whole life.
  
  
• Claude Code: The same agentic engine, but living in the terminal for developers. This is where all of this started, and it’s still the power tool of the family but if you’re not a dev, just stick with Cowork.
  
  
• Agentic chat features: Even regular Claude chat can now search the web, run code, create files, and connect to your tools (Google Drive, calendars, and so on) via those MCP connections we just covered.

ChatGPT

OpenAI folded its old Operator and Deep Research tools into one thing: Agent mode.

Flip it on in any conversation and ChatGPT gets its own virtual computer. It can browse websites, fill in forms, run code, and analyse files.

It’s session-based rather than always-on. You kick off a task, it churns away for 5 – 30 minutes, and it pauses to ask permission before doing anything consequential (like spending your money, thankfully).

Microsoft Copilot and Google Gemini

Not to be left out:

• Copilot can now take multi-step actions inside Word, Excel and PowerPoint like building that pivot table or restructuring the document rather than politely suggesting how you might. Microsoft has even launched its own Copilot Cowork (yes, they borrowed the name, and yes, it’s powered by Anthropic’s Claude under the hood).
  
• Gemini has its own agentic mode plus “Gems” which are custom mini-assistants you can shape for specific tasks and a growing library of pre-built agents. Access to this depends on your license.

So what does this mean for you?

Two things.

First, before you spend a penny on some shiny “AI agent platform” a vendor is flogging you, check what’s already sitting in your existing AI subscription. There’s a decent chance the agentic capability you need is one tab or toggle away.

Second, this changes the skill you need to build. With assistants, the skill was prompting. With these built-in agentic features, the skill is delegation. You need to describe an outcome clearly, decide what access to grant, and know when to check the work.

That second one matters more than any tool. These features are genuinely useful, but they’re also confident enough to be wrong at speed. Treat them like a capable new team member: brief them well, give them access to only what they need, and review the output before it goes anywhere important.

The hype merchants will tell you agents replace your judgment. In reality, they raise the price of not having any.

To Agent or not to Agent, that is the question

Every tool has its time and place.

I say that too often. Much like LLMs, and AI in general, Agents aren’t the answer to everything. Knowing when (and when not) to call upon the powers of an AI agent is a skill in itself.

My best advice is actually stolen from an engineer at Anthropic (creator of Claude).

Barry Zheng (Applied AI team at Anthropic) gave what I class as a legendary answer to the growing trend of people trying to apply agents to every problem, even when simpler systems would suffice.

“Don’t go after a fly with a Bazooka”

Barry Zheng (Applied AI team at Anthropic)

Magnificent!

I see this so much these days with a lot of tech.

So many tasks can be done in a few minutes by a human, but we’ll spend hours trying to get AI to do it. Surely, that’s counterintuitive to the goal?

Barry also shared this useful slide from one of his live talks (if you’re reading this, Barry, I’m not stalking you – promise!).

And to echo what Quentin Villard shared on LinkedIn, here’s a quick framework to figure out the best tool for the job:

• If a task requires interacting with external services or your digital environment and is not set up as a workflow or agent, you need to do it yourself. Use a degree of common sense here. If the task is simple or you enjoy it, use that supercomputer in your head, aka the brain.
• Choose an AI workflow for repeatable, rule-based tasks where you want predictable automation.
• Choose an AI agent for tasks where you have a goal and want the AI to dynamically figure out the steps, acting as a flexible assistant.

How to build and experiment with AI agents today

Tier 1: Already inside your AI subscription (zero extra cost)

If you pay for a mainstream AI assistant, you already own an agent builder. Start here.

• Claude Projects + Skills (Anthropic): Bundle instructions, knowledge files, and reusable “skills” so Claude behaves like a specialist for a specific job. Combine with Cowork on desktop for agentic execution. Included in Claude Pro ($20/month).
  
• Custom GPTs (OpenAI): Configure a reusable agent in ~10 minutes: write instructions, upload knowledge files, toggle on tools like web browsing and code execution. No code, built into ChatGPT Plus. Pair with Agent mode for tasks that need autonomous browsing and file creation.
  
• Copilot Agent Builder (Microsoft): If your organisation lives in Microsoft 365, Copilot includes a built-in agent platform with pre-built agents and a builder for your own, with permissions and compliance handled automatically.

Best for: Your first agent.

Tier 2: No-code agent platforms

For agents that need to act across your other tools like email, calendar, CRM, forms.

• Zapier Agents: The lowest-friction option if you already automate anything. Zapier has evolved from “if this then that” automation into an agent builder where agents make decisions and act across apps, with 8,000+ integrations. Free plan available; AI agent plans from $50/month.
  
• Lindy: Built specifically for non-technical users. Drag-and-drop workflow builder, ready-made templates, 4,000+ integrations, and agents that can work together (one qualifies leads, another sends follow-ups). Good for inbox triage, scheduling, and follow-up style agents.
  
• Gumloop: Agent builder with a built-in AI assistant (“Gummie”) that builds agents for you from a description, premium models included without needing your own API keys. From $37/month with a free plan.
  
• MindStudio: Strong balance of ease and flexibility, with access to multiple AI models rather than locking you into one vendor. Does need a little bit of more technical expertise than others on this list.
  
• Make: Visual, flowchart-style automation with AI steps. Slightly steeper learning curve than Zapier but cheap and appeals to visual thinkers.

👀 Watch usage-based pricing. Many platforms charge per task run or credit, so costs scale with adoption, which means, a £15 experiment can become a £150/month habit.

Final thoughts

Of course, there’s much more to say about agents.

But for 95% of the humble humans in this world, this is what you need to know.

This space will continue to grow faster than my cups of tea can brew, but that doesn’t mean you need to be flying at the same speed.

Deep and meaningful understanding requires a moment or two to breathe.

Agents are here, they’re useful, and it will only become easier to access them in shared marketplaces.

As a bonus, here’s a few more resources to shape your knowledge:

• AI Agents explained for beginners
• How to evaluate the actual worth of AI agents in your work
• The truth behind the AI Agent hype with Boston Consulting Group (BCG)
• How to build effective AI agents (Anthropic)

Go forth, human.

→ If you’ve found this helpful, please consider sharing it wherever you hang out online, tag me in and share your thoughts.

Before you go… 👋

If you like my writing and think “Hey, I’d like to hear more of what this guy has to say” then you’re in luck.

You can join me every Tuesday morning for more tools, templates and insights for the modern L&D pro in my weekly newsletter.