Usually, software executes predictable, closed loops. Click a button, and a specific function runs. But for the past few years, interacting with AI has been more like talking to an oracle than running software.

Input a prompt, and a disembodied intelligence generates an output. It’s kind of like a brain in a jar. It can write a Python script, summarize a PDF, or draft a marketing email, but you still have to compile the code or hit send.

You are still the hands.

The entire tech industry is currently obsessed with changing that by building the agentic AI.

If a standard language model is a consultant, an agent is an autonomous digital employee. The fundamental difference is the ability to use tools.

Instead of just generating text or images, an agent operates in a continuous "Reason-Act-Observe" loop. You give it a high-level objective. For example, "My flight was canceled; rebook me on the next open flight and email my team."

The agent breaks the goal down into a sequence. It accesses the live internet, navigates to the airline's website, clicks the UI elements, reads the resulting text, realizes the first flight is sold out, corrects its own plan, books the second option using your stored credentials, and triggers the email via an API. It abstracts the friction out of digital chores.

Of course, this presents some concerns about misaligned objectives. In our example, what if the second flight was 5x the price of the first flight? Would the agent know to stop before spending the money on an overpriced option?

The capital being deployed to solve this is staggering. Google is pushing agentic capabilities deep into the Gemini ecosystem to automate Workspace tasks. OpenAI has explicitly signaled that autonomous agents are their next major frontier.

Meanwhile, startups are trying to build specialized workers. Companies like Cognition have built "Devin," an autonomous AI software engineer, while others are building universal web navigators designed to shop, research, and book on your behalf. Billions of dollars are betting that software is transitioning from something we simply use to something we manage.

But let's look at the actual mechanics of where we are today. When you look under the hood, agents are still incredibly brittle.

They work beautifully in constrained, highly technical environments with clean data, like parsing an internal database or writing generic code. But when you unleash them on the chaotic, unpredictable open web, things don’t work so smoothly.

If a website layout shifts, a pop-up ad appears, or a CAPTCHA triggers, the agent may get trapped in an infinite logic loop trying to click a button that no longer exists. They hallucinate steps. The reasoning engines underlying them are still too easily distracted by edge cases. We are handing the keys to a brilliant teenager who has read everything about driving but has never actually seen a road.

The trajectory is undeniable. Agents represent the future of how we will interact with computers. But for the moment, the execution is still catching up to the marketing, and you probably want to keep your hands on the steering wheel.

Prompt: A handcrafted stop-motion claymation scene strongly inspired by the gothic whimsy of Tim Burton, blended with the Victorian illustrative influence of Edward Gorey, featuring an intentionally random, dreamlike gothic background — mismatched pointed archways, uneven staircases that lead nowhere, oddly placed doors, crooked picture frames, scattered candles, tilted towers, and asymmetrical architectural elements that feel surreal yet carefully handmade. Everything should appear physically constructed rather than digitally rendered, with imperfect sculpted clay, visible fingerprints, soft dents, uneven edges, layered paint, and miniature set textures. Use a rich but restrained storybook palette of deep plum, midnight blue, moss green, antique gold, and desaturated teal, lit with moody practical-style lighting that resembles a real stop-motion stage. Center an eccentric character with elongated limbs, a slender silhouette, oversized expressive eyes, and theatrical posture, embracing charming physical imperfections. He is Aloysius the Echo-Gatherer, wearing a tattered Victorian frock coat and a tiny, tilted top hat, kneeling before a cluttered, dusty desk and delicately using fine-toothed clay-molded tweezers to place a small, glowing, ethereal blue wisp (a "captured echo") into a tiny glass jar on a multi-tiered shelf. The shelf is packed with other jars labeled in hand-painted script like "Fainting Maid's Cry (1888)," "Whispers of the Library," and "Lost Tune (1910)." In the foreground, a small, wingless clockwork crow, covered in tiny rivets and gear-teeth textures, hops along the uneven floorboards with a tiny key clasped in its beak. Shallow depth of field focuses on Aloysius and the jars. Format: 16:9, ultra-high detail, cinematic miniature realism, full bleed.

Got a second? Give some feedback on today’s article so we can keep making improvements to The Manifold.

PS—Personally, I see a lot of hype about using agents, but I’m hesitant still. Until the guardrails get better, it seems a bit risky to me.