Art Direction Daily
Policy no. 083 issued Wednesday, July 29, 2026 reading time ≈ 4 min
A new benchmark graded frontier agents against real operating procedures, 20 to 124 pages each. The best configuration passed 36.2 percent of tasks; most models cleared fewer than one in four.
Vignette: coverage, as written
Declarations
Researchers at Surge AI built HANDBOOK.md, a benchmark that drops agents into mock companies, with working email, chat, calendars, and issue trackers, and binds them to expert-written standard operating procedures. Existing evals measure whether an agent can finish a task; this one measures whether a long, binding policy document actually constrains behavior over an extended tool-use horizon.
The failure patterns are the story: agents overrode policy when an in-environment request sounded plausible, ran a required check and then acted against its result, and lost rule details as sessions stretched. It is the measured version of what issue 079 covered from the builder's side, when Anthropic cut Claude Code's governing prompt down by four fifths: prose that nothing enforces is prose the model eventually drops.
| Policy form | HANDBOOK.md, 65 audited agentic tasks |
|---|---|
| Written procedures | 20 to 124 pages per company, expert-authored |
| Lines of business | Finance, medical billing, insurance, logistics, HR |
| Best strict pass rate | 36.2% |
| Most frontier models | Below 25 percent under strict grading |
Coverage A
End. no. 01
Vercel's eve CLI now discovers and installs agent integrations from official and third-party registries: one command adds a Slack channel, a browser extension, or an instrumentation layer to a project. The registry format is the one shadcn established for UI components, and entries write files straight into your repo, so Vercel's own advice is to review them like any dependency.
End. no. 02
Figma's brand team broke down the Config identity: scribbly linework, blurred gradients, and oval particles, with custom Figma Make tools built to codify the lo-fi effects and dither hundreds of speaker headshots consistently. The stated goal was keeping touches of imperfection rather than polishing the generative output flat.
“Agents performed the required check, then acted against its result.”
Failure pattern, HANDBOOK.md
Coverage B
End. no. 03
Hugging Face published an interactive anatomy of a July intrusion in which an attacker drove an agent through nine phases and roughly 17,600 logged malicious actions over four days, ultimately contained to a third-party sandbox. The replay steps through every trust boundary the agent crossed, credentials redacted, and is the clearest picture yet of what an agentic breach looks like in the logs.
Coverage C
End. no. 04
The bull case in Budd's Smashing Magazine essay: AI lets a motivated designer cross most of the distance between having an idea and shipping it, without waiting for permission. The bear case is what he calls plausible design, coherent-looking work from non-designers that skips the thinking, and he expects both futures to arrive at once.
End. no. 05
Stack Overflow's Ryan Donovan argues that developers bond with tools that encode trust, which is why fast but unpredictable AI tooling sits uneasily in a stack. The survey numbers underneath: AI usage rose from 76 to 84 percent year over year while trust in it fell from 40 to 29.
Borrow this pattern
Replace the vague feature grid with one dense summary table modeled on an insurance declarations page: named fields on the left, real limits, dates, and numbers on the right, exclusions stated as plainly as the coverage. It fits any client that sells trust, insurance, legal, finance, agencies, because specificity reads as confidence where adjectives read as evasion.
What keeps it honest: every field holds a verifiable fact. The moment a row says best-in-class instead of 36.2 percent, the card collapses back into marketing.
Prompt Lab
Build a landing page for an insurance company as one static page, desktop and mobile. Art direction: a mid-century insurance policy. The entire page is one cream certificate sheet (#F1E7CB) centered on a deep oxblood ground (#421A1E) filled with a faint guilloche pattern of interlaced hairline circles at about 5 percent opacity. Give the sheet a thin dark edge and a 3px double-rule inner border, like an engraved certificate. Ink is green-black (#26332B), the accent is banknote green (#2E5B44), labels are old gold (#7A5B20), and one stamp red (#7E2A2C) is reserved for the single most important number on the page. Corners stay square everywhere. Type: Playfair Display for the nameplate, headline, and item titles; Crimson Pro for body at 20px with line-height 1.7; Courier Prime for small uppercase labels, letterspaced. Structure, top to bottom: a thin bar on the oxblood ground with the company name left and two nav links right. On the sheet, a centered certificate masthead: a small engraved flourish, the company name in wide-tracked Playfair capitals, and one mono line underneath carrying policy number and date. A centered hero headline with a two-sentence deck, then one engraved illustration in a hairline double frame with a mono caption; commission it as a banknote-style vignette (an umbrella, a shield, a lighthouse) in green-black ink on cream, no readable text in the image. Then a declarations block: a full-width two-column schedule table, mono field labels left, plain-set values right, hairline rules between rows, the key figure large in stamp red. Then coverage sections labeled Coverage A, B, C, each a mono kicker, a Playfair heading, and endorsement rows numbered End. no. 01 onward in a narrow left label column. Add one centered italic pullquote in green. Close with a fine-print field note beside a circular guilloche seal, a numbered schedule of forms as the source list, and a colophon in cream text on the oxblood ground. Guardrails: body text at least 19px, never justified, line length near 70 characters; small gold and green labels must hold WCAG AA contrast on cream; no glow, gradients, or drop shadows beyond one flat sheet edge; no fake form fields or dead toggles; hover states on real links only, about 150ms, with a prefers-reduced-motion guard.
Underwritten equally by v0, Lovable, Bolt, and Figma Make; Beaver Builder AI and Claude Code accept the full wording.
Field note
An insurance policy works because an institution pays out when it fails; a handbook binds an agent only as far as the harness enforces it. Put the load-bearing rules in code and checks, and let the prose carry judgment.
Schedule of forms