Taste-bound tasks and agent selection (2026-06-12)

The same model, on the same day, audited a whole vault cleanly and then needed roughly twelve rounds to write one acceptable hero line. The gap looked like capability. It wasn’t — it was a task-type difference that agent workflows should be designed around.

The diagnosis

Two kinds of task got conflated. The wide work (restructure the standards, backfill decisions, recover the Atlas paper trail) runs on reasoning over readable things: a repo, a git log, written rules. The tagline was a taste-matching task — the spec existed only in Wedge’s head, and the agent’s only access was strikes. Wide tasks reward generation; taste tasks punish it.

The failure mode has a name: generating instead of eliciting. Twelve rounds of candidates, each absorbing the previous strike, each missing the next unstated rule. The collapse came when Wedge wrote four candidate lines himself and the agent explained why they worked — one round from there to acceptance. The solving information existed the whole time; the agent never asked for it.

Looking beats guessing, every time it was tried. Both mid-arc breakthroughs came from artifacts, not iteration: reading the actual dictionary entries (the 休 form line carried more register information than five rounds of feedback) and fetching the live site (which settled a layout dispute in one look). Wedge’s own words, mid-session: “when you look at it, generally you understand better.”

The rules extracted

• Taste-bound tasks start from exemplars. Two examples the owner likes, or a pointer at the real thing, before any candidates are generated. This is now standing procedure in agent memory; it should be standing procedure for any agent (Fable, Opus, Grok) given front-facing words.
• Strikes are training data, not waste. The twelve rejections became the Taglines Are Catalog Lines section and the nouns-counts-states rule — the law exists because the failure happened in front of the standard-setter. The cost was real; so was the artifact.
• Model selection stays empirical. The open question Wedge raised: whether Opus serves better on narrow middle-tier tasks. Tier rankings say no; observed behavior is the better instrument. The honest position is to A/B it on real tasks rather than argue from the spec sheet — and the elicit-first procedure probably matters more than the model picked.

Ruled out

• Reading the twelve rounds as a model-capability verdict. The failure was procedural (generate-first on a taste task), and the fix is workflow, not tier.

Outstanding

• Apply the exemplar-first rule to the next front-facing task across any agent and see whether the round count actually drops — that’s the falsifiable test.
• The Opus comparison, if Wedge wants it: same narrow task, both models, count the rounds.