Coding agents need your intent more than your code

The viral example. A GitHub repository called "reverse engineer anything" has become the latest proof, for some, that AI agents can replace engineers. The video makes the opposite point. The repo works because the target application already exists, so the agent has an unambiguous reference and can validate every step against it. That is a powerful demo, and it is not evidence that agents can build anything. Most real work does not come with a finished example attached.

The actual bottleneck. The speaker has not written code in over a year, but spends more time planning and validating than before. His software factory takes a PRD, splits it into tasks, and builds an application autonomously. The hard part is communicating intent: specifying what good looks like, choosing references, and building a harness that asks questions up front and compares output at the end. This is engineering the harness instead of the codebase.

Where it breaks. Repeated attempts with strong models and a detailed PRD produced a 2D game when the speaker wanted 3D, because he could not describe the visual style. The breakthrough came only when he used a Legend of Zelda game as a graphics reference and added a workflow that compared each result. It is a nice demonstration, but it is one person's hobby project, not production data, and the "weird circle" glitch shows the validation loop is still rough.

A useful test. The video offers a practical filter: if the next model is twice as good, does a skill still matter? Prompt engineering and hand-tuned rule files matter less; planning, validation, sub-agents, hooks, and the overall agent workflow matter more. That framing beats the usual "learn to code" debate and is the strongest idea here. It also implies engineers who enjoy coding mechanics may find the job less rewarding.

What is missing. The video is optimistic about the individual builder, but silent on teams. It does not address code review, architecture decisions, on-call, security, staffing, or how an organization decides what agents should build. The Firecrawl segment is a sponsored pitch and adds nothing to the argument. Evidence elsewhere is anecdotal, self-reported, and lacking metrics. The central observation, that an agent without a reference writes good code that misses the point, is credible but not new.

The takeaway. The actionable signal for engineering leaders: invest in people who can define problems, build references, and design agent workflows, not just in more capable models. The video overstates how universal the harness-builder role will become, but it is directionally right. Most of the work around coding is becoming work around intent, and teams should spend engineering effort there.