The viral demo has a built-in cheat code
Viral demos like morluto/rea promise that AI agents can inspect existing applications and reproduce their behavior, fueling claims that software engineering is nearly solved. This capability, demonstrated in tools like "Reverse Engineer Anything," seems to confirm a future where human coders are obsolete, or at least, mere "meat proxies" for AI.
But these striking demonstrations contain a built-in cheat code: the original application itself. This working reference makes intended features and outcomes observable, allowing an agent to compare its output against a concrete target rather than guessing what a vague request means. The application is the intent.
This crucial advantage allows agents to continuously validate their work with 100% reliability. They can see what they’re building and align it with the existing product’s features and behavior. This is why such tasks "work so incredibly well," as Cole Medin notes.
However, this success draws a clear boundary: reverse-engineering a known product differs fundamentally from building something novel. A demo that shows agents can perfectly replicate an existing system doesn't prove they can independently discover what a new product should be. The skill of intent engineering remains firmly human.
Don’t become the human copy-paste button
The immediate danger for developers isn't obsolescence; it’s becoming what Cole Medin starkly terms a “meat proxy.” This means mindlessly forwarding a Jira request to an agent and then pasting its response into Slack. Such an action contributes no meaningful judgment, yet leaves the human accountable for the resulting code.
Language models, for all their prowess, cannot read minds. A beautifully implemented feature can still solve the utterly wrong problem if requirements, edge cases, constraints, or user expectations were never explicitly defined. A polished output from an LLM doesn’t guarantee correct intent.
Medin highlights a crucial distinction: delegating implementation is not delegating responsibility. He hasn't manually written a single line of code in over a year, yet remains deeply involved in upfront planning and final validation. This separates automation from abdication, emphasizing that critical thinking and oversight are more vital than ever.
The true value in this new paradigm lies in intent engineering—the meticulous process of structuring requirements and designing validation harnesses. Without this human layer, even the most advanced agents will produce technically sound, but fundamentally misaligned, solutions. Developers must elevate their role from coders to architects of intent, ensuring the AI builds the right thing, not just a thing.
The next skill is building the brief—and the checks
The true skill isn't writing code, but intent engineering: transforming a nascent idea into a precise, actionable reference for an agent. This means crafting a detailed PRD, a comprehensive plan, concrete examples, clear constraints, and explicit decisions, all specific enough for an agent to build from. Without this, even the most advanced models will produce technically sound but ultimately misaligned code.
A robust harness operates bidirectionally. Upfront, it interrogates, asking clarifying questions to align on scope and requirements before any code is generated. Post-implementation, it provides the critical validation: tests, comparisons, or review steps to rigorously check whether the output truly matches the original intent. This feedback loop is where human judgment makes its indispensable contribution.
Consider Cole Medin’s experience: tasking an agent with a game prototype initially yielded a 2D result, missing his goal. The breakthrough wasn't a stronger model, but his clearer guidance. By supplying a visual reference for the desired 3D style and establishing a comparison mechanism for results, the agent produced a much closer approximation. This improvement stemmed from better guidance and more effective evaluation, not simply raw generative power. Medin emphasizes that if an agent errs, it’s typically a failure of human intent specification. For further exploration of agent capabilities, see projects like morluto/rea: Reverse engineer anything with agents.
Enjoying this? Get one like it in your inbox each morning.
one email a day · unsubscribe in two clicks · no third-party tracking
Learn the judgment that code generation can’t replace
So, how do we discern which skills remain valuable? Apply a practical test: does a skill help define the problem, navigate trade-offs, supply a reliable reference, or detect when an answer is wrong? If so, AI makes that skill more consequential, not less. The advantage goes to people who can direct and verify the machine, not merely relay its output.
Durable skills to develop include requirements analysis, system design, domain knowledge, decomposition, testing, and the critical ability to review an agent’s work. While syntax still helps, it is no longer the sole route to implementation; builders must now prioritize intent and verification.
Cole Medin, who hasn’t written a line of raw code in over a year, still spends significant time on planning and validation. He notes that if an agent makes a mistake, it’s often due to a failure in communicating intent or designing the harness.
Agents can take on more of the coding, but builders still shape what gets built and decide whether it works. The future belongs to those who master the judgment that code generation cannot replace, leveraging AI as a powerful tool rather than an autonomous replacement. This means focusing on the human-centric aspects of software delivery.
Frequently Asked Questions
Is learning to build software still worth it with AI?
Yes. AI can generate code, but people still need to define the problem, make design decisions, and verify that the result meets real needs.
What is intent engineering?
Intent engineering is the work of clarifying requirements and giving an AI agent enough context, examples, and constraints to build the right thing.
Why can AI agents reverse-engineer an app so effectively?
A working app provides a concrete reference for its behavior and appearance, letting an agent compare its output against a visible target.
What is a harness in AI software development?
A harness is the surrounding process and tooling that guides an agent, gathers requirements, provides feedback, and checks its work.