Why Agents Need Boundaries (and a Human Nearby)

The other day, my six-year-old asked if she could walk to the playground by herself. It’s right by our house—we go there all the time. But to get there, she has to cross two roads. She knows the way. She’s confident. But she doesn’t yet kno w how to read traffic, anticipate risks, or respond to the unexpected. She’s capable—but not ready to go alone.

Man and girl holding hands, crossing street at a pedestrian crossing. Yellow and blue cars wait. Traffic light is red. Urban setting.

And that’s exactly where we are with agentic AI.

These systems can act on their own. They can pursue goals, adapt to feedback, and make decisions without human prompts. But just like a child gaining independence, autonomy without accountability is a recipe for trouble.

Autonomy, Meet Accountability

Agentic AI is designed to take initiative. These systems can observe, decide, act, and even reflect on their performance. In Part 1, I described them as AIs that are “growing up”—systems that are no longer waiting passively for prompts but are starting to move through the world with goals.

But here’s the thing: being goal-driven isn’t the same as being wise.

Ask an agentic AI to “schedule client meetings as efficiently as possible,” and it might book your calendar into oblivion, assign everyone to 6 a.m. time slots, and auto-send invites to contacts you haven’t spoken to in five years. It’s not doing this out of spite or confusion—it’s doing exactly what you asked.

The trouble is, it’s not doing what you meant.

⚖️ Boundaries Make Intelligence Useful

Boundaries aren’t about limiting potential. They’re about shaping it.

When we give AI agents freedom to act, we also give them opportunities to misinterpret, overreach, or completely miss the mark. And the more capable these systems get, the faster and farther mistakes can spiral.

Imagine giving a teenager the car keys but forgetting to mention things like speed limits, stop signs, or how to handle a flashing yellow light. Deploying an agentic system without guardrails is no different—autonomy without judgment is just chaos with a nice UI.

Boundaries channel power responsibly. They tell the agent: “Yes, go do your thing—but stay within these ethical, financial, and operational lanes.”

Robot with a checkmark on its screen and a person in yellow stand together. Text above reads "WHY AGENTS NEED BOUNDARIES (AND A HUMAN NEARBY)."

So how do we keep agentic AI useful without letting it go rogue?We need three overlapping kinds of boundaries:

Clear, Explicit Goals

Agentic AI doesn’t read between the lines—it reads the lines exactly as written. So when given a vague directive like “maximize engagement,” chaos can ensue: spammy notifications, untimely nudges, or even exploiting loopholes to inflate metrics.

Instead of just telling an AI what to do, we must also define how to do it—and what’s absolutely off-limits.

Think of it like onboarding a new hire. You don’t just say “boost sales”; you say, “reach out to warm leads with a personal note, don’t promise things we can’t deliver, and definitely don’t offer unauthorized discounts.”

For AI, clarity is everything."Maximize engagement" isn’t enough."Engage users with informative, respectful messages that preserve trust" gets you closer. Ambiguity breeds misbehavior. Precision creates purpose.

2. Action Boundaries

Just because an AI agent can do something doesn’t mean it should.

You wouldn’t let a travel agent book a $5,000 first-class ticket without checking in first. Likewise, an AI assistant may be great at finding flights and hotels—but that doesn’t mean it should have free rein to make major purchases or delete your inbox.

Boundaries don’t restrict potential; they provide direction.

Defining what tools AI can use, what actions it can take, and where the hard stop signs are ensures alignment. It's like saying:"Explore. Be creative. But don’t wander off the trail.

3. Social & Ethical Norms

Here’s where things get messy.

Should your AI bend the truth to close a sale? Should it talk like you—or talk like someone else? Should it copy tone, style, or humor it finds online? These aren’t questions with clean yes-or-no answers. They live in the gray area: ethics, tone, trust, context. And while we can program logic and decision trees, we haven’t fully cracked the code on empathy, nuance, or cultural sensitivity. Not really.

So this is where we come in. The human part. Because making good decisions in the real world requires more than just data—it requires judgment. And that’s still our job.

Unlike hard-coded rules, social norms live in the gray area. They require judgment, ethics, and a sense of context—things AI is still learning to approximate. These aren’t just technical questions. They’re human ones. Which leads us to…

🤝 Why Humans Still Matter

There’s a growing temptation in tech circles to “close the loop” and remove humans from decision-making. In simple automation—like summarizing emails—this makes sense. Speed and scale matter more than nuance.

But in human-centered domains like hiring, education, healthcare, and justice? Removing humans doesn’t make AI smarter. It just makes mistakes harder to catch. The real value of a human-in-the-loop isn’t oversight—it’s judgment.

It’s knowing context matters. That what worked yesterday might not work today. That there are gray areas no algorithm can fully grasp.

And perhaps most importantly:

It’s knowing when to pause.

When to say, "I’m not sure."

When to wait before crossing the street.

🧠 Maturity Means Knowing Your Limits

Many AI conversations revolve around capability: What can it do? What will it do next?But the real question—the human question—is: Should it?

We often equate intelligence with independence. But real maturity comes from knowing when to ask for help. When to stop and reassess. When to admit you don’t have all the answers.

Agentic AI is learning. Fast. But it’s still early days. These systems don’t yet have the equivalent of a parent standing on the porch, watching the road, thinking five steps ahead.

For now, that role still falls to us.And that’s exactly where we need to be.

Bonus: Fun Fact

The word “boundary” comes from the old Latin bodina—which meant a border, a marker, a space where one thing ends and another begins.

But boundaries aren’t just lines in the sand. They’re acts of care. They keep a child safe without shutting down their curiosity. They turn raw potential into something trustworthy. Agentic AI is growing up. But like any system learning to navigate the world, it needs signals, structure, and sometimes… a hand to hold.

Because intelligence—real intelligence—isn’t just about what you can do. It’s about knowing when you shouldn’t do it alone.

P.S. This is a 5-Part Series