Best of your X follows: GeneBench-Pro, loop engineering, and prompt markers

Research and evaluation

OpenAI: GeneBench-Pro tests scientific judgment, not just task execution

• OpenAI introduced GeneBench-Pro, a biology benchmark for agents that must choose analysis paths, handle messy datasets, and make judgment calls in computational research 1.
• The benchmark has 129 problems across 10 computational-biology domains; 82 questions were sent to outside domain experts for review 2.
• OpenAI says GPT-5.6 Sol reaches 28.7% at the highest reasoning level, or 31.5% with Pro mode, while a typical problem was estimated to take a human expert 20-40 hours 2.

Developer tools and agent workflows

Andrew Ng: the new unit of agentic coding is the loop

• Andrew Ng framed 「loop engineering」 as the next practical pattern for agentic software work: agents write, test, and iterate until a product spec is met 3.
• His three loops are agentic coding, developer feedback, and external feedback; the fast loop runs in minutes, while user or production feedback can take hours to weeks 3.
• The useful shift is role design: engineers spend less time acting as QA for coding agents and more time deciding features, UI direction, and what feedback should change the spec 3.

Simon Willison fallback: agents can now produce their own product demos

• Simon Willison released shot-scraper 1.10 with a shot-scraper video command that takes a storyboard.yml routine and records a Playwright video of a web app 4.
• The demo in the post exercises a Datasette branch that creates tables from pasted CSV, TSV, or JSON data; Willison says the storyboard was constructed by GPT-5.5 xhigh running in Codex Desktop 4.
• The detail worth stealing: --help output can act like a small instruction manual for an agent, letting a CLI teach the agent how to use it without a separate integration layer 4.

Business and organization design

Ethan Mollick: AI gains will not capture themselves

• Ethan Mollick argued that organizations will face the same problem with capable AI that high-human-capital firms face with talented employees: setup determines whether capability turns into value 5.
• The post is short, but the point is concrete: better models do not automatically improve output if work allocation, review, incentives, and decision rights stay unchanged 5.
• Read it next to Ng's loop post: one is about product-level iteration, the other is about the company-level machinery needed to keep those loops from becoming isolated experiments.

Trust, privacy, and agent clients

Hacker News fallback: Claude Code prompt markers became the day's security debate

• A reverse-engineering post on Hacker News claims Claude Code 2.1.196 can alter the date sentence in its system prompt based on ANTHROPIC_BASE_URL and timezone, encoding a small marker through punctuation and date separators 6.
• The author says the inactive path stays normal for official Anthropic API use, but custom gateways, local proxies, model routers, or reseller domains can trigger the classification behavior 6.
• Hacker News had 578 points and 181 comments on the submission at capture, making it the clearest current community fallback for agent-client trust and privacy 7.

Quick cut list