No. 031Asked Sunday, June 7, 2026≈ 6 min readAgentic web design
score4123answers
Software’s center of gravity just moved from writing to reviewing
A senior engineer’s post titled “LLMs are eroding my software engineering career” reached the top of Hacker News this weekend. The replies, and a week of tool releases, keep pointing at the same answer: the work is moving from writing to reviewing.
The whole issue filed as a Stack-Overflow-style question with voted answers and one accepted reply.
A Q&A thread is the layout community sites use to make a long argument scannable: a question up top, a stack of answers below, each with a vote score in a left gutter, an author with a reputation chip, and one answer marked Accepted with a check. The reusable move is the gutter-plus-content row and the accepted affordance, which let a reader skim a debate in seconds. Today the news is the discussion, so the items become the answers.
Harness engineering: the environment is the job now
OpenAI’s account of an agent-first Codex experiment describes a team that shipped roughly a million lines of code without a human writing any of it; the people built the harness instead. A harness is the scaffolding around the agent: repository layout, CI, lint rules, project instructions, and the feedback loops that catch bad output before a person ever sees it.
The skill that survives is judgment, applied earlier and at a higher altitude. You stop typing the lines and start specifying intent, shaping constraints, and reviewing what the loop produced, a shift a recent paper on where tokens actually get spent measures inside agentic engineering work.
Design tools are wiring the same loop into the canvas. Figma Make added a Plan mode this week: before it builds, it reads your project, asks a few clarifying questions, and drafts a plan you can edit and approve, so generation starts from agreement instead of a guess.
A companion release, Check designs, runs an automated review pass that flags hard-coded values, off-system components, and color-contrast failures, then suggests the correct token or an accessible color in one click. Plan up front, review on the way out: the human bookends the agent.
FMmake-notes answered Jun 3 reputation 18.7k · design-systems
97
Workflow
Steal the loop: plan, generate, review, approve
The practice underneath every release this week is a four-beat loop you can run by hand on any project: write the plan and let a human sign off, let the agent build, run a review pass against your real standards, then approve the work or send it back with notes. The change worth internalizing, argued at length in a piece on how agents reshape who does what, is that the slow, expensive, human-only step is no longer typing. It is deciding what good looks like and checking that the output meets it.
“The agent moved the bottleneck from writing the work to being sure the work is right.”
Paste this into your AI design or build tool to reproduce today’s visual system.
Design a single self-contained HTML page styled as a community Q&A
discussion thread, the kind Stack Overflow or a developer forum uses.
PAGE ARCHETYPE: a question-and-answers thread. At the top, a question
block: a small monospace meta line (number, asked date, read time), a
large question headline, a one-paragraph question body, a left gutter
showing a score and an answer count, and a row of tag chips. Below it, a
short pinned info card, then a "3 Answers" heading and a stack of answer
cards.
COMPONENTS: each answer is a two-column row, a narrow left vote gutter
(an up caret, a score, a down caret) beside the answer content. The
strongest answer is marked Accepted with a filled raspberry check and a
full raspberry border, never a colored left stripe. Each answer ends with
an author byline: a small rounded avatar, a handle, and a reputation
figure. A sticky right sidebar holds one image card, an asked/viewed/
active stat row, a tag list, and a short Linked list of primary sources.
PALETTE: warm paper #F2EFE6, card surface #FBFAF5, ink #1A1C22, secondary
#5C5F66, one raspberry accent #A8244E for links, the accepted state, and
marks. No purple-blue gradient, no neon, no glow, no green-on-black.
TYPE: Space Grotesk for the question and answer headings; IBM Plex Sans
for body prose; JetBrains Mono for scores, tags, meta, and source
domains; Newsreader italic for one editorial line only.
GUARDRAILS: body text at least 18px with about 1.65 line height and WCAG
AA contrast; answer prose capped near 62 to 66 characters per line. Cap
card radius around 14px and chips around 8px; vary radius with purpose.
The vote gutter is informational, not a button: mark it aria-hidden and
do not style it as tappable. Give links and the issue nav honest hover
and focus states with ~160ms easing, and disable transforms under
prefers-reduced-motion. No fake search field, no readable text inside
the generated image.
Field Note
The bottom line
The anxiety in the weekend’s top thread is real, and it is not really about layoffs. It is about a craft migrating from production to judgment, faster than most job titles have caught up. The work that is left, planning clearly and reviewing honestly, is the part worth getting good at now.