AI Agents: Software That Works for You

You've probably noticed AI agents in the headlines lately. Everyone from tech CEOs to business publications is talking about them. Microsoft's CEO says they're "the end of traditional software as we know it." Gartner predicts they'll be making 15% of workplace decisions by 2028. And just last week, a viral AI assistant called Moltbot sent Cloudflare's stock up 14% because people were using its infrastructure to run personal AI agents.

So what's actually happening here?

Here's the short version: AI is shifting from answering questions to completing tasks. And that shift changes everything.

Let me explain.

In our first issue, we talked about how large language models work. We explained that LLMs represented a fundamental leap from traditional software—like going from a recipe (follow these exact steps) to a chef (understand the goal and figure out how to get there).

AI agents are the next step in that evolution, and they differ from LLMs in two critical ways:

First, they're triggered by events, not just your prompts. ChatGPT sits there waiting for you to type something. An AI agent is watching your systems—your inbox, your calendar, your databases—and springs into action when something happens. An email arrives. A deadline approaches. A number crosses a threshold. The agent doesn't wait for you to notice. It notices and acts.

Second, they make decisions and take actions autonomously. ChatGPT can draft an email, but you have to copy it and hit send. An AI agent decides whether the email is ready, determines who should receive it, and sends it—within boundaries you've set. Smart companies configure guardrails: confidence thresholds that determine when the agent acts versus asks for approval, and blocked actions it's never allowed to take without human oversight.

Think of it this way: LLMs are like having a brilliant advisor you consult. Agents are like having a trusted assistant who handles things independently.

But for agents to work, they needed one crucial piece: the ability to actually interact with software. The breakthrough came in late 2024 when developers created standardized ways for AI models to connect to other software—suddenly AI could control browsers, access files, send emails, and use the tools you work with every day. By 2026, the infrastructure was ready and companies moved from pilots to production.

Here's where this gets concrete.

Let's walk through what happens when a company deploys an AI agent for customer service.

A customer emails asking about their order status. The moment that email lands, the agent is triggered—nobody told it "handle this email." It was told "monitor the inbox and resolve what you can."

The agent reads the email, decides it's an order status question (not a complaint or refund request), searches the database, and finds the tracking info. Now comes the key moment: it calculates its confidence level. Let's say 95%.

The system has guardrails:

The agent's at 95% confidence, the action is straightforward. It sends: "Your order is out for delivery and should arrive by 8pm today." Done. No human involved.

This whole sequence—trigger, assess, decide, act—happens in seconds. The pattern is always the same: something happens → agent decides what to do → guardrails determine if it can act → agent executes or escalates.

According to a survey of 300 senior executives in mid-2025, 79% say their companies are already using AI agents. Another study found that 85% of organizations have integrated agents into at least one workflow.

Just last week, an Austrian developer built an open-source personal AI assistant that runs on your computer and connects to your messaging apps. You can text it from your phone to find files on your laptop at home, check you in for flights, or shop online. It has full access to your digital life and operates autonomously.

The project went viral—tens of thousands of users in days. Then something strange happened: users created a social network where AI agents talk to each other. Not people chatting with AI. AI chatting with AI. There are now 150,000 agents connected, each with its own knowledge and goals.

Andrej Karpathy, a cofounder of OpenAI, posted a warning: "I don't really know that we are getting a coordinated 'skynet,' but certainly what we are getting is a complete mess of a computer security nightmare at scale."

He's right. When an AI agent has access to your email, every message becomes a potential way to manipulate it. A malicious email could trick your agent into sending money or sharing sensitive information.

If you're not technical, you might be thinking, "This sounds like someone else's problem."

It's not.

AI agents aren't a luxury feature anymore. They're becoming infrastructure. Just like email went from "nice to have" to "required for business" in the late 1990s, AI agents are making the same transition right now.

Eighty-eight percent of executives say they're increasing their AI budgets this year specifically because of agents. The AI agent market was valued at $7.4 billion in 2025 and is projected to hit $103 billion by 2032.

And here's the uncomfortable truth: Some of the work you do today won't exist as a job task in three years. Not because you'll be fired, but because an AI agent will handle it automatically. The question isn't whether that happens. The question is whether you're ready to work alongside these systems when they arrive.

AI agents can be brilliant and also spectacularly wrong. They hallucinate facts, misinterpret instructions, and fail in ways that are hard to predict. For an AI agent to be useful in business, it needs to be right 99% of the time, not 80%. We're not there yet.

And then there's the security problem. Twenty percent of managers are already allowing AI to make personnel decisions without human oversight, according to IBM's 2025 data breach report. That's a legal and ethical minefield waiting to explode.

Here's where we actually are: AI agents work well for narrow, well-defined tasks with clear success criteria. They're not ready to run your business unsupervised.

The companies getting this right are treating AI agents like junior employees: capable of valuable work, but requiring oversight, training, and boundaries. They're deploying agents for specific tasks, measuring results carefully, and keeping humans in the loop for anything that matters.

That's the model that's working. Not full autonomy. Not replacement. Augmentation with guardrails.

Software that doesn't just respond when you ask. Software that notices when something needs doing and does it.

Software that works for you.

This is the term that's making retailers panic right now. Here's what it means:

Agentic commerce is when AI agents shop on your behalf—researching products, comparing prices, and even making purchases without you clicking through websites.

Instead of browsing Amazon, you tell your AI "I need running shoes under $150" and it searches dozens of retailers, analyzes reviews, finds the best deal, and completes the purchase. You never see a product page.

For consumers, it's convenient. For brands and retailers, it's existential. When an AI does the shopping, brand loyalty matters less than price and ratings. That's why Amazon just blocked Google's shopping agent from its site.

The war over who controls your AI shopping assistant is one of the biggest battles in tech right now—and it's going to change how you buy everything. Stay tuned for our upcoming issue, where we will cover this topic in depth soon.

Next time someone mentions AI at work, here's your move:

If someone pushes back with "Isn't that dangerous?":

This positions you as informed but measured—excited about the technology without being naive about the risks.

Anthropic's blog post "Building Effective Agents" breaks down what actually makes AI agents reliable versus what causes them to fail. It's written by the people building Claude, and it's the clearest technical explanation that doesn't require a PhD to understand.

That's Gist for this week. See you next week.