Curious Companion: AI Agents Explained (For Non-Coders)

LISTEN TO THE EPISODE

In this episode I break down what AI agents actually are and how they differ from regular chat tools. Before diving in, I run through the recent flood of new model releases from Anthropic, OpenAI, and xAI, including the full story on the Fable 5 suspension and reinstatement. The episode ends with a practical mini cheatsheet on when an agent is actually the right tool for your workflow.

It’s Everywhere

Today we are talking about a term that is everywhere in the AI world: Agents.

Real quick, before we jump into agents, I want to run through all of the new models that were recently released, both to just keep you in the loop and also to set the stage because these new models are being developed in order to improve agentic capabilities.

So let’s talk about the eleventy billion releases that went down in the past two weeks or so.

Everybody’s Still Competing

Anthropic, OpenAI, and xAI all released new models within days of each other, a clear reminder that these companies continue to be competing in the self-created race to AGI (artificial general intelligence).

Anthropic: Fable 5 and Mythos 5 finally came back online after being suspended on June 12th (check out episode 48 for the full rundown of Fable and Mythos). A quick recap of what happened with the suspension and reinstating:

Fable launched June 9th.
Three days later, Amazon’s security researchers found a jailbreak, a way to trick Fable into ignoring its own safety rules, that could get the model to identify software vulnerabilities and in at least one case produce actual exploit code, the kind someone could use to break into a system.
The US Commerce Department ordered Anthropic to cut off access for any foreign national anywhere in the world. Since Anthropic had no way to verify a user’s nationality in real time, they shut both models down globally rather than risk violating the order.
For the next two weeks, Anthropic worked with the government and with Amazon to build a new safety classifier targeting that jailbreak, and NIST’s AI safety center tested the fix and signed off on it.
The Commerce Department lifted the restriction on June 30th (Mythos had already started coming back to a small group of vetted organizations a few days before that), and Fable returned to everyone July 1st.

Anthropic maintained the whole time that the models were safe and the response was overblown, but they fixed the issue anyway, and gave affected users 50% extra weekly usage through July 12th to make up for the outage.

Separately, also on June 30th, Anthropic released Claude Sonnet 5, which is now the default model for anyone on Claude’s free or Pro plan. If you use Claude and haven’t specifically upgraded to a higher tier, you’re already using it whether you noticed the switch or not.

OpenAI: released GPT-5.6, which comes in three versions, Sol, Terra, and Luna, going from most powerful down to fastest and cheapest. Artificial Analysis has GPT-5.6 Sol scoring 59 on its Intelligence Index, one point behind Fable 5’s 60, while costing roughly a third as much per task, making it arguably the best pick specifically for coding work, at a meaningfully lower price. Expect to see more back and forth battling between OpenAI’s Sol and Anthropic’s Fable. Alongside the models, OpenAI also launched a separate product called ChatGPT Work, an actual agent, not just a chatbot, that drafts documents, spreadsheets, and decks across your connected apps.

xAI: released Grok 4.5, their first model built specifically for coding and agent style work, trained partly on real session data from Cursor, the coding tool SpaceX bought for something like $60 billion back in June (I maintain that this acquisition is actually hugely problematic but no one is talking about it so I shall just continue to watch from the sidelines). Whack ass Musk called the model “Opus-class” and said it’s roughly comparable Opus 4.7. That’s Musk comparing his own model to Anthropic’s, not something anyone outside xAI has independently verified, so feel free to ignore it completely. Artificial Analysis, a third party benchmarking outfit, ranked Grok 4.5 fourth overall on its Intelligence Index, behind Fable 5, GPT-5.5, and Opus 4.8, though still ahead of every open-weight model and all of Google’s Gemini lineup. Where it did lead the pack was agentic tool use specifically, posting the single best score of any model tested on that measure, and using a fraction of the tokens Opus 4.8 needed to do comparable work.

Does any of this change what you personally experience? If you’re paying $20 a month for one of these tools, mostly no. The benchmark charts these companies publish measure things like how well a model fixes bugs in a codebase, or how it holds up across a long multi-step task in a terminal, categories that matter a ton if you’re a developer running these models at scale, and barely at all if you’re asking one of these tools to help you write an email or think through a decision. For everyday use, all four of these releases are more capable than what came before them, and they’re closer to each other in quality than the headlines make it sound. The real gap shows up in specialized, technical work, not in a regular conversation.

So…WTF is an Agent?

Every single one of those releases we just discussed is more “agentic” than the model it replaced, meaning the model itself is better at planning multiple steps and using tools with less hand holding. But none of those four releases IS an agent.

An agent is a computer program. It is an actual piece of software built around a model (Sonnet 5, GPT-5.6 Sol, Fable 5, etc). It gives the model three things:

Tools it can call. A search function, a code interpreter, a browser it can click around in, access to your files or other apps.
A loop. It doesn’t take one action and stop and wait on you. It keeps cycling on its own: do a step, look at what happened, decide what to do next, do that step, repeat, all without you sending a new message each time just to keep it moving.
A way to check its own work. It looks at the result of each step before deciding what to do next.

The model is the part that makes the decisions. The agent is the program wrapped around it that turns those decisions into actions, tracking where the task stands, and continuing until the model decides that the job is done, or it hits a stopping point you set.

How this differs from a regular chat: in chat, you ask, it answers, you go do the thing yourself. With an agent, it plans, acts, checks its own work, and keeps moving without you approving each individual step. It doesn’t matter if you’re watching it happen live or you walk away and check back later. The autonomy between checkpoints is what makes it an agent.

“Building an Agent”

When people say “I built an agent” it can indicate wildly different amounts of work depending on the context:

Low effort: Configuring something that already exists. Giving a product like a Custom GPT or a Claude project some instructions, reference files, and turning a few tools on, basically the same process I walked through in the Custom GPT episode. No code involved, you’re just telling existing software what to do and what it can touch.
High effort: Actually writing the program. Using something like the Claude Agent SDK or a framework like LangChain, where a developer decides exactly which tools it gets, how many steps it can take, and what counts as finished.

Both get called “building an agent” but they are not the same job.

“Spinning Up” or “Running Multiple Agents”

This usually means someone started multiple instances of an agent program and handed each one a task. Picture five copies of the same program, each working a different piece of a problem at the same time, sometimes with one lead agent splitting up the work and pulling the results back together at the end.

It’s not five different AI models having a meeting. It’s the same model, running several times in parallel, the way you might have five browser tabs open instead of one.

Wait, Is Regular Claude an Agent?

The short answer: not quite.

An example to help elucidate things: I had Claude (specifically Sonnet 5) help me do research for this episode. I had it search the internet multiple times, read the results, decide what to search for next based on what it found, write summaries, and check its own work. Claude was able to plan, act, check, and decide the next step, without me telling it which individual search to run. That’s absolutely agentic behavior.

So is that an agent? Not quite. The autonomy is capped at a single response. All of that searching and deciding happened inside one reply to me. Once it answered, it stopped. It doesn’t keep working after that, it doesn’t own the task across time, and it’s not deciding what to do next until I send another message. Claude Code or Cowork are built to keep running across a longer stretch, sometimes without me checking in after every step.

Thus, the real line of agent vs non-agent isn’t “can it do things”, it’s scope of autonomy. Does the system only work within one exchange, even a complicated one, or is it built to own a task across time with less need for me to show up at every checkpoint? Regular Claude gives you a taste of agentic behavior inside a chat wrapper. A dedicated agent product is that same capability inside a different kind of wrapper, one built for longer, less supervised work instead of a single back and forth, running longer, with more independence.

But Doesn’t Claude Code Ask Permission?

If you’ve been in the computer you’ve had Claude Code hit you with the permissions questions. One might then ask, If it’s asking me what to do, is it really acting on its own?

In this case Claude is not asking you what to do, it’s asking if it’s ok to do what it’s already determined to be the next step.

“What should I do next?” means the human is forming the plan. The system hit a wall and handed the actual decision back to you. That’s regular chat behavior, even if it happens in the middle of a task.
“Can I run this command?” means the agent already decided what it wants to do, it formed the plan and chose the next step, and it’s only pausing to get a yes on something risky or hard to undo, like running a command, deleting a file, or editing something live. You’re not the planner in that exchange, you’re a gate. The agent is still steering.

Between two permission checks, the agent might read a bunch of files, weigh a few different approaches, and land on a plan, all without you involved at all. That never happens in regular chat, where you’re re-engaged after literally every single step because there’s no ongoing plan for it to work from between your messages.

Human-in-the-Loop

That pattern, an agent that pauses for approval on risky moves, has an actual name in the industry: human-in-the-loop. This is the dominant way agents are being used right now, and is more common than fully autonomous.

This approach is adjustable, not fixed. You can loosen something like Claude Code’s permissions to auto-approve certain actions, or run it fully autonomous with Claude never checking in with you at all (the command for this is “dangerously-skip-permissions,” which should tell you something about how Anthropic feels about making that the default setting).

Permission-asking isn’t proof something is “less of an agent.” It’s a safety setting sitting on top of a capable agent.

You’re Probably Already Using One

Agents are just a concept. They’re already sitting inside tools you have access to and might already be using:

Claude Code handles software work end to end: building features, fixing bugs, testing and cleaning up code, and running whatever’s needed to actually ship it
Claude Cowork handles knowledge work end to end: research, analysis, writing, spreadsheets, presentations, pulling from your connected tools and files and delivering a finished result
ChatGPT Work drafts documents, spreadsheets, and decks across your connected apps
Claude in Chrome browses and takes actions on websites for you
Grok Build and Cursor are coding specific agents

For each of these, you’re handing off a task to a program that acts on its own to complete the assignment, not asking a question to a model that can only provide you with an answer.

Who’s Actually Using This?

Per that list above, if you use Claude Code or Claude Cowork, you’re already using an agent. Same goes for ChatGPT Work or Claude in Chrome. So this agent discussion isn’t “nobody’s using this yet”, it’s “here’s the actual name for something some of you are already doing, and here’s what it means for the rest of you.”

The enterprise (large companies/big corporations) numbers get thrown around a lot too, so for context: McKinsey found most companies are still just experimenting with agents rather than running them at real scale, and that even among companies claiming they’ve adopted agents, only a small fraction have rolled them out broadly. Adoption is absolutely happening right now outside of the tech world, it’s just still in its early days.

Do You Need an Agent?

AI continues to be “here’s a solution, go find a problem”, and agents are no exception. Depending on the use case, they can genuinely be helpful, so I put together a mini cheatsheet.

Good fit for an agent right now:

Repetitive, well defined tasks you already know how to do yourself (let the robots do the work)
Tasks where you’re totally fine reviewing the output before anything actually goes live, which, if you’ve been listening, is exactly how I run my own podcast automation
Multi-step research or busywork where you don’t mind waiting for a background result instead of babysitting it in real time

Not there yet, proceed with caution:

Anything you can’t easily verify
High stakes stuff like sending something, spending money, or publishing, without a review step in between
If you don’t yet trust regular chat outputs without double checking them, handing an agent five unsupervised steps probably isn’t the place you start building that trust

Most of us don’t need to become agent power users tomorrow, but it is, IMO, worth understanding what they are so that we can identify when they can be put to work.

How I Used AI This Week

Each episode I share a quick example of how I, or someone I know, used AI that week.

This time, I’m sharing about Tom, husband to my volleyball friend, Diana. We recently went to Switzerland for a volleyball tournament, and Tom happily tagged along. Tom is way more of a hiker than a beach volleyball spectator, so he used ChatGPT to plan out an entire day trip to a location about three hours away from where the tournament was located.

He put in what he wanted to do, his time frame, and his budget, and it created an entire itinerary for him. Not too shabby.

Da Wrap-up

Agents aren’t some far-off future thing, they’re already sitting inside tools you might already be using, and every new model release is being built with agentic capabilities in mind. At its core, an agent is just a program that plans, acts, checks its own work, and keeps moving without you holding its hand at every step. Whether or not you need one right now depends entirely on your use case, but it’s worth knowing what they are so you can actually spot when they’d be useful.

As always, endlessly grateful for you and your curiosity.

Catch you next Thursday.

Maestro out.