You are now the whole team

A common claim is that AI makes software engineering easier, and that anyone can now code. The first half is true: writing code was slow and expensive, and an agent genuinely removes that cost. The second half is true only for software nobody depends on. This series argues the opposite of the usual conclusion. We were always the engineer and the bricklayer at once, because in our trade putting the walls up is itself an ultra-qualified job and there was never a cheaper pair of hands to give it to — so the engineering always competed with the building, and lost. Now that something else can lay the bricks, one person is left holding every role a team used to spread around: product owner, project manager, architect, platform engineer, QA and reviewer. Every claim here is measured against two real repositories, and the conversations are quoted on both sides.

9 of 9 parts published.

Parts

  1. At long last, only an engineer

    We were always the engineer and the bricklayer, because in software putting the walls up is itself an ultra-qualified job. Something else can lay the bricks now — which does not make you less of an engineer, it makes you only an engineer.

  2. The terminal I could not build

    The thing I wanted for years and could not afford to build, what it sounds like when a terminal is navigable by heading, and why I still did not start by asking an agent to write it.

  3. The gap between what you said and what you meant

    Extrapolation is the whole reason you use an agent, and the whole risk of using one, because they are the same mechanism. So the question is never how to stop it — it is where it should be wide and where it must be narrow.

  4. The night that produced no code

    Five hundred and forty lines of documentation, six files, and not one line of code — then fourteen more commits before the first one. Six jobs got done that evening, and I did not notice I was switching between them.

  5. Conditions, not instructions

    What I actually typed on the first night, and why almost none of it was an instruction. Five moves, quoted from the record — including the ones where I told the agent I did not understand a word of what it had just proposed.

  6. An hour a day

    What the method from last time actually produced, measured rather than asserted: where the hours went, how much had to be done twice, and the afternoon a machine that had never built any of this was ready in nine minutes.

  7. I stopped building the product to build the tool

    Twenty-six days between one commit on the terminal and the next, because every pull request cost me an evening of listening. What I built instead, and the four questions that separate this from procrastination.

  8. The server is everywhere, a bridge is somewhere

    The architecture conversation that filled the twenty-six days: one reader-agnostic server, one bridge per screen reader in whatever language that reader dictates, and a decision to reject Rust for a reason I have never seen written down.

  9. Six things that got through

    Every mechanism this series has described, and the six things that beat them anyway — the wrong remedy for a right diagnosis, the machine that never complains, the gate that lied about itself, and the rule I was proudest of, whose hole an agent found and was entirely right about.