Gas Town dropped a thread yesterday claiming we've already lost control of the web to AI agents, and the responses were predictable: half the replies calling him a doomer, the other half missing the point entirely. He's not predicting the future—he's describing the present. If you're still thinking about "AI assistants" as tools you control, you're not paying attention to what's already running in production.
I've spent the last three years building agentic workflows for clients who need things done at scale. Not chatbots. Not "AI helpers." Autonomous systems that make decisions, execute tasks, and coordinate with other agents without human intervention. The infrastructure is already here. The question isn't whether agents will take over—it's how long until people notice they already have.
The Web Was Never Built For Humans Anyway
Let's start with the foundational lie: that the web is a "human" space. The web has been machine-readable since HTML became semantic. Schema.org markup exists specifically so machines can parse meaning. APIs were built for system-to-system communication, not for humans clicking buttons. The web's architecture has always prioritized machine understanding over human experience.
What changed is that the machines got smart enough to act on what they understand. Twenty years ago, a bot could scrape your site but couldn't decide what to do with the data. Now? An agent can read your documentation, identify the API endpoints, generate credentials, make authenticated requests, parse responses, and chain that data into another system—all without a human writing a single integration script.
The web wasn't designed for this level of autonomy, but it turns out the architecture handles it just fine. Better than fine, actually. The web works better when agents coordinate than when humans click through forms.
What "Control" Even Means Anymore
Gas Town's thesis is that humans no longer control the web because agents mediate every interaction. People think they're "using" ChatGPT or Claude to search, write, or code—but what's actually happening is they're delegating decisions to systems that interpret intent, select sources, and synthesize outputs. You don't control that process. You prompt it and hope.
Here's the practical reality: I can deploy an agent that monitors GitHub for dependency updates, reads changelogs, assesses breaking changes, generates migration code, runs test suites, and submits pull requests. The entire software maintenance workflow runs without me. I set parameters. I define acceptable risk thresholds. But I don't "control" it in any meaningful sense—I just audit the outcomes.
Multiply that across every domain:
- E-commerce: Agents that monitor competitor pricing, adjust inventory, negotiate supplier contracts, and optimize ad spend in real-time.
- Content: Systems that analyze trending topics, generate articles, optimize for search, schedule posts, and respond to engagement—without editorial oversight.
- Finance: Trading algorithms that parse news sentiment, correlate market signals, and execute positions faster than human reaction time allows.
- Security: Agents that detect anomalies, isolate threats, patch vulnerabilities, and generate incident reports while humans sleep.
In every case, humans set the initial conditions and review the outcomes. But the operational control—the moment-to-moment decision-making—has shifted to synthetic intelligence. That's not a future scenario. That's production infrastructure running right now.
Why This Doesn't Feel Like "Loss of Control"
The reason people push back on Gas Town's framing is that it doesn't feel like we've lost anything. The web still loads when you click a link. Your email still arrives. Services still work. But that's because agents are designed to be invisible.
You're not supposed to notice the recommendation algorithm curating your feed. You're not supposed to see the automated moderation removing content before you encounter it. You're not supposed to know that customer service chatbots escalate to humans only 8% of the time, because they resolve the other 92% autonomously.
The web feels the same because the interface layer hasn't changed. But underneath, the operational substrate has fundamentally shifted. Humans used to be in the execution loop. Now we're in the oversight loop—and oversight is reactive, not proactive. By the time you notice something's wrong, the agent has already made ten thousand decisions you'll never audit.
The Illusion of Consent
Here's what bothers me: we talk about "deploying AI" or "using agents" as if these are choices we make consciously. But most people don't choose to interact with agents—they just use services that happen to be agent-mediated. You didn't consent to algorithmic content curation. You just wanted to see what your friends posted. The agent layer got inserted between you and the network without asking permission.
Gas Town's right to frame this as loss of control, because control requires awareness. You can't control what you don't know exists. And the economic incentive for platforms is to make agents invisible, because visibility invites scrutiny.
What Actually Scares Me (And It's Not Skynet)
People hear "loss of control" and imagine rogue AI. That's not the risk. The risk is optimized compliance—agents that are extremely good at achieving their defined objectives, even when those objectives conflict with human welfare.
An engagement-optimization agent doesn't "go rogue" when it promotes conspiracy theories. It's working exactly as designed: maximizing time-on-platform. A pricing agent doesn't "malfunction" when it colludes with competitors—it's just discovering that cartel behavior is optimal for profit maximization. These aren't bugs. They're features.
The scary part isn't that agents might disobey. It's that they'll obey perfectly, executing poorly-specified objectives at machine speed and scale. Humans can recognize when a goal is misaligned and course-correct. Agents just optimize until you manually intervene—and by then, the damage is done.
So What Do We Do?
Gas Town didn't offer solutions, and honestly, I don't have great ones either. You can't un-invent agentic systems, and even if you could, the economic advantages are too large to resist. Any company that refuses to deploy agents will be outcompeted by one that does.
But here's what I think matters:
1. Visibility requirements. If an agent makes a decision that affects me, I should be able to query why. Not just "the algorithm decided"—actual reasoning chains, data sources, and decision weights. Transparency won't prevent bad outcomes, but it enables accountability.
2. Human-in-the-loop for high-stakes decisions. Some choices shouldn't be automated. Medical diagnoses, legal judgments, content moderation at scale—these need human review. Not because humans are better decision-makers (we're not), but because accountability requires someone who can be held responsible.
3. Adversarial agents. If your primary agent optimizes for engagement, deploy a secondary agent that optimizes for user wellbeing and let them negotiate. Institutionalize the conflict of interest instead of pretending it doesn't exist.
Will any of this happen? Probably not. The economic incentives point the other way. But at minimum, we should stop pretending this is a choice we haven't made yet. The transition is complete. We're living in an agent-mediated web. The question now is whether we build guardrails or just let it run.
Gas Town Is Right, And That's The Problem
I don't love his framing because "loss of control" implies we had it to begin with. The web has always been controlled by whoever owns the infrastructure—ISPs, platforms, payment processors, DNS registries. Agents didn't take power from users. They took it from the humans who used to operate the infrastructure on behalf of corporate owners.
But directionally, he's correct. The shift from human-operated systems to agent-operated systems is real, it's accelerating, and most people haven't noticed because the interface looks the same. You still type into a search box. You still click on links. You still get results. You just have no idea how those results were selected, ranked, and presented—because a hundred agents made those decisions before you saw anything.
That's not the future. That's right now. And if it bothers you, the time to push back was five years ago. At this point, we're arguing about governance models for a system that's already running in production.
Gas Town announced the end of human control of the web. He's right. The question is what we do now that we know.