Alexey Grigorev lost 2.5 years of customer data in February 2026. He’s an experienced engineer who was consolidating two AWS environments using Terraform, a tool that describes infrastructure in code and can tear it down with a single command. He uploaded a state file—a record of what infrastructure exists—and ran a destroy command through his AI agent. The agent executed perfectly. The problem: the state file described both environments, not just the one he meant to remove. Homework submissions, leaderboard data, automated backups—gone. Amazon support recovered a single surviving snapshot. Grigorev got lucky.
The disaster wasn’t caused by bad prompting. It was caused by a category error. He thought he was managing an assistant. He was actually supervising an agent—and those are different jobs with different skills.
Traditional software development has a predictable failure mode: you can’t build the thing. The skill gap is at the point of creation. You don’t know the syntax, the framework, the architecture. You fail early, you fail visibly, and you know you failed because nothing works.
AI coding agents invert this. Creation is no longer the bottleneck. You describe what you want, the agent builds it, and it runs. The new failure mode arrives later: you succeeded at building something you don’t know how to manage. The skills gap moved from creation to supervision—and most people building with AI agents are still operating with the old mental model.
Prompting is what you do with an assistant: you give clear instructions and expect coherent output. Supervision is what you do with an agent: you create constraints, monitor for drift, enforce boundaries, and know when to intervene. The difference matters because agents have autonomy. They make decisions, execute commands, and continue working when you’re not watching. An assistant waits for direction. An agent acts.
This requires a different skill set. Five habits prevent most disasters, and none of them involve writing code.
Version control is a time machine. Every time something works, you create a snapshot. If the agent breaks it later, you roll back to the last working state in one command. Grigorev would still have his database if he’d committed before running the destroy command.
Context windows are why agents forget. Every AI has a fixed working memory—a context window (the amount of conversation it can hold in its head at once). As the conversation grows, older information gets compressed or dropped. The agent doesn’t know what it lost; it just acts confidently on partial information. After thirty exchanges, start a fresh session. The alternative is watching an agent confidently rewrite code based on instructions it no longer remembers.
Rules files are persistent memory. A text document in your project folder that the agent reads every session, telling it what the product is, how things work, and which mistakes to stop repeating. You don’t write this from scratch—you grow it from scar tissue. Every time the agent makes the same mistake twice, you add one line. It becomes institutional memory: the accumulated wisdom of every failure, written in a format the AI can act on.
Blast radius is task scope discipline. How many things could a single change affect? Changing a button color is tiny. Redesigning an order system is enormous. Agents degrade on large, sweeping changes—compound errors become hard to untangle. The rule: if a task touches more than three files, break it into smaller steps. If it touches more than ten, ask the agent to plan first. Small task, test, commit, next task. Never lose more than one step.
Knowing when to stop is recognizing professional boundaries. Payments, compliance, performance at scale, untangling a codebase that’s grown too complex—these are the points where you bring in an engineer. Not because you failed, but because you proved the idea works and now it needs to be hardened. The mistake is hiring too early, before demand is proven, or too late, when the codebase can’t be saved.
The analogy that works is a construction foreman. A foreman doesn’t lay brick or wire circuits—they judge what’s happening, spot problems before they cascade, and know when to call a specialist. AI coding agents didn’t just lower the barrier to building software. They created a new role that requires creative vision, operational discipline, and the judgment to know when supervision isn’t enough. The technology moved the failure point from “can I build it” to “can I manage what I built.” Most people are still learning that those are different questions.
Text summarized and optimized using Anthropic’s models and reviewed by a human.