Autonomous AI agents create more work than they remove.

Autonomous AI agents are currently sold like relief. They are supposed to take work off your plate. Delegate the task. Connect the tools. Add the MCP servers. Write the perfect system prompt. Give it a browser, terminal, GitHub, Slack, email, calendars, docs, search, memory, and the blessed little permission slip to act like a junior employee with root access.

Then the agent starts working, and somehow you now have two jobs: the original task, and supervising the artificial intern trying to solve it by opening seventeen drawers, inventing three assumptions, and calling that a plan.

This is the part the demos politely step around. In a clean demo, the agent has a narrow goal, a friendly environment, obvious success criteria, and no messy human context. Real work is not like that. Real work is a swamp of half-remembered decisions, implicit priorities, political landmines, old constraints, weird dependencies, and the kind of "obvious" information nobody writes down because everyone involved already knows it. Except the agent. Obviously. The agent knows nothing. It has vibes and a tool belt.

And yes, you can make the tool belt beautiful. You can build the best MCP setup. You can expose the right APIs. You can write system prompts with sections, rituals, guardrails, examples, punishments, prayers, and tiny laminated commandments. You can give it memory. You can give it file access. You can give it every integration short of a forklift license.

It still does not understand the situation the way a human does.

That is the core problem. Not intelligence in the abstract. Not whether it can summarize, code, search, refactor, draft, classify, or run commands. It can do all of that, sometimes well. The problem is judgment across context. The problem is knowing which connection matters, which detail is decorative, which instruction is stale, which shortcut is dangerous, and which apparently small thing will become a tedious little disaster by Thursday afternoon.

Humans are annoying, biased, tired, emotional, and regularly powered by caffeine and spite. But humans are also very good at situational awareness. A decent human worker can hear a request and immediately factor in the invisible stuff: who asked, why now, what happened last time, which system is fragile, which stakeholder says "quick question" before detonating a week, and which solution is technically correct but socially unusable.

Agents do not have that. They approximate it. They simulate it. They write sentences that look like they have it. But the moment the task requires a real sense of the surrounding mess, they start producing homework for you.

They ask for clarification when they should inspect. They inspect when they should decide. They decide when they should stop. They make progress in the most literal, trackable, dashboard-friendly sense while quietly increasing the amount of cleanup required later. It is motion wearing a fake mustache and calling itself execution.

The babysitting is not incidental. It is the product experience. You watch the agent because you cannot fully trust its prioritization, its sense of risk, its interpretation of vague goals, or its ability to notice when the best next step is "do not touch that." So you hover. You correct. You redirect. You tighten the prompt. You remove a tool. You add a tool. You explain the same obvious business rule again, like feeding coins into a haunted vending machine that dispenses pull requests.

After a while, the workload changes shape instead of shrinking. You do less typing, maybe. But you do more reviewing. More scoping. More containment. More checking whether the agent changed the one file it was supposed to change or also refactored a sidebar because it got bored and found a pattern. Congratulations, you have automated keystrokes and invented management.

This is especially frustrating because the ceiling is real. That is what makes it maddening. Agents can be genuinely useful. They can search a codebase faster than you. They can draft boring boilerplate. They can run repetitive checks. They can stitch together information, generate first passes, and act as a tireless assistant for bounded tasks with clear verification. That is not nothing. That is useful. I am annoyed, not a candlelit monk yelling at electricity.

But "useful assistant" is not the same thing as "autonomous worker." The industry keeps trying to sell the second thing by showing the first thing in a nicer jacket. A tool that helps me move faster is valuable. A tool that requires me to become its nervous project manager while pretending to replace project management is, frankly, taking the piss.

The better mental model is not a genius employee. It is a very fast, very literal contractor with memory issues, uneven taste, and no real skin in the game. Give that contractor a small, well-defined job with a visible finish line and they might save your afternoon. Give them a vague business problem, a pile of connected tools, and the instruction "figure it out," and you deserve whatever circus arrives in the diff.

This is also why better prompts only get you so far. Prompting can shape behavior, but it cannot magically install lived context. It cannot give the model the quiet, embodied sense of "this smells wrong" that comes from experience inside a team, a system, or a customer problem. It can mimic caution. It can list risks. It can quote your rules back at you with the obedience of a cursed parrot. But the underlying assessment is still brittle.

The gap is not tool access. The gap is interpretation.

That is why autonomous agents often feel worse the more powerful they get. A weak agent fails safely because it cannot do much. A powerful agent with broad access can create impressive amounts of plausible damage. It can move files, open issues, edit docs, send messages, deploy changes, and generally produce the kind of confident activity that looks productive until a human has to understand what the hell happened.

There is a future where this gets better. Models will improve. Tooling will improve. Evaluation will improve. Maybe agents will become better at preserving context, asking fewer stupid questions, noticing risk, and respecting the sacred ancient art of not making things worse. Lovely. I hope so. I would enjoy being wrong here. Being disappointed by software is not a lifestyle I chose; it is one the industry keeps shipping to me.

But right now, the honest version is simple: autonomous AI agents are not relief. They are leverage with supervision costs. Sometimes the leverage is worth it. Sometimes it absolutely is not. If the task is bounded, visible, reversible, and easy to verify, let the goblin run. If the task depends on messy human context, unstated priorities, careful tradeoffs, or actual situational judgment, keep your hand on the leash.

The fantasy is that agents remove work. The reality is that they often transform work into babysitting, reviewing, correcting, and cleaning up after a machine that can operate tools faster than it can understand why those tools matter.

That is not autonomy. That is a Roomba with sudo.