One agent, 119 blog heroes, and the scaffolding that made them shippable

Start with the money, because it is the weakest part of the case. Giri puts the avoided image-generation bill at $5 to $20 for the first batch, plus charges every time he wanted something redone ^[2]. Across 119 posts that works out to between four and seventeen cents an image ^[22]. Nobody assembles a three-layer graphics system to save seventeen cents a picture, and the account is better for admitting it: Giri's other stated reason was that generated bitmaps would look as generic as the design he was replacing ^[14].

What the system bought instead was determinism. The illustrations are plain JavaScript and SVG in three layers: a component library of characters and office props, 16 composition functions, and a specification file that assigns a scene and annotations to each post ^[3]. An SVG turbulence filter adds a seeded wobble, so the output looks drawn by hand while the geometry underneath stays predictable ^[17]. Sixteen compositions covering 119 posts means each scene is doing about seven posts of work [c20b]. The unit of production is therefore a function, not an image. Recoloring the character, which was adapted from the MIT-licensed Xiaohei project and turned yellow ^[16], is one edit followed by a re-render through a headless browser at 1600x900 ^[18]. In a prompt-per-image workflow it is 119 rolls of the dice.

The failures are where the template gets useful. Three are on the record: a category badge rendered black on black ^[4], a coral topic color chosen after the agent had been told to avoid red ^[5], and a development server that had to be switched to no-cache because Giri's browser kept showing him the old build ^[6]. Read them as a list of missing checks rather than as charming stumbles. Black on black is an absent contrast test. Coral is proof that a stated constraint is not an enforced one, which is why "no red anywhere" ^[11] had to be repeated by a human rather than asserted once. The stale build is the worst of the three, because for a while the reviewer and the artifact had diverged, and no approval gate survives that.

The gate that did work was structural. The agent produced two complete themes behind a switchable CSS flag, one bookish and one modeled on an editorial magazine ^[12], which turned a matter of taste into a fork with two shippable branches. Giri picked the magazine, then directed typography and palette from there ^[12], and the redesign entered production through CI rather than by hand ^[1]. That is the shape worth copying: constraint stated up front, volume delegated as components, judgment spent once at a defined choice, promotion through the same pipeline everything else uses.

There is a reflexive joke in the setup. Munder Difflin exists because running several coding-agent terminals at once made Giri the message bus, conflict resolver and long-term memory between them ^[20], and the product wraps command-line agents with shared memory, mailboxes and isolated worktrees to automate that job [c19b]. The blog redesign, the contained test of the thesis, ran as one agent with one human catching color mistakes ^[8]. The coordination layer is still the person, and on this evidence that is not yet optional.