Thoughts on Coding Agents (rakyll)

A short essay from rakyll.org, 3 July 2026, marking a shift the author dates to December 2025. The framing: models are being trained to produce trajectories that work for people who already know where they want to go. What the model sells is a shorter distance between a clear goal and a working artifact, which makes the interesting question "how cheap is it now to get from a decision to a working artifact?"

That reframing decides who benefits. Not the person handing an agent a vague goal and hoping, but the person with a strong internal model of the outcome, who can steer, correct course, and recognize a good trajectory on sight. The agent amplifies direction; it does not supply it. If you know where you are going, it takes you there faster.

The name was wrong

The second observation is that "coding agent" turned out to be a bad description of the boundary. We named them for writing software and quietly assumed that was the extent of their usefulness. But generating code is not a narrow skill — code is the most general interface anyone has for acting on the world. It calls any API, transforms data, drives tools, and automates almost anything a person can describe. Once a model could turn intent into working code reliably, it inherited the reach of everything code touches. The coding part was almost incidental.

The bottleneck moved

The strongest section is about organizations. Inside a large company, the real cost of building something that solves a hard problem was rarely the typing. It was inventing something worth building, and then coordination: organizing people, aligning teams, convincing stakeholders while the code was being written. Execution overhead was dwarfed by social overhead, and you spent your best energy earning permission to start, then earning it again to continue.

Coding agents remove that permission system. An individual can go from intent to a working prototype without assembling a group first, which lowers the bar for showing something real instead of arguing for it. The essay does not treat this as unambiguously good. The same collapse in execution cost makes cookie-licking easier, and makes it trivial to fork an idea or a project when nobody is empowered to make a broad decision — chaos for organizations that depend on someone being in charge. Small high-trust teams are thriving; permission-bound bureaucracies are struggling. That, the author argues, is the actual explanation for why the picture looks so uneven right now.

The worry

The closing concern is transmission. Senior engineers are no longer spending enough time educating newcomers, because the amplifier rewards people who already know where they are going and the payoff for teaching has dropped. But knowing where to go is learned by watching someone do it. If that stops being passed on, the result is a generation that can execute and cannot set the goal, and losing the ability to decide where to point all this leverage is the risk the author wants kept in view.

Connections

The transmission worry lands in the middle of an existing cluster: dont-outsource-learning on the individual version of the same trade, cognitive-debt on what it costs downstream, and skill-atrophy-supervision-paradox on the circularity where the skills needed to supervise agents are the ones agent use erodes. rakyll adds the organizational mechanism — not that individuals choose to stop learning, but that the incentive to teach evaporates.

"Amplifier of direction, not a substitute for it" is the same boundary vibe-engineering draws between keeping the decisions and delegating them, and the same reason short-leash-ai-method is written for experts only. individual-engineer-agency is the pre-agent baseline this essay is measuring against: the claim there is that a single engineer controls much less of the product than they assume, and rakyll's argument is that agents move that line.

who-manages-the-agents covers the adjacent political claim, that the point of all this should be making the median worker more capable rather than concentrating capability. Whether the productivity story holds up at all is measuring-ai-coding-productivity — this essay offers no measurement and does not pretend to.