Working Surfaceissue · 21 Nov
Issue · 3 links
Issue

Saturday
21 November 2026

OpenAI, Nielsen Norman Group and the speakers at a Sutter Hill Ventures event describe AI agents doing investigation, design and drafting work. In each account, people decide what the agents read and approve what reaches customers.

Swipe for 02 to 03
21 Nov
The note

Three accounts show AI agents investigating, designing and drafting, while people choose what the agents read and decide which results go out. Each team also built a way to learn that an agent is wrong before customers do.

Okay Zed and Cyrus Forbes, engineers at OpenAI (21 Nov · 01), describe how the company keeps ChatGPT fast. AI agents now help ship code so quickly that engineers could no longer investigate every slowdown by hand. The team built the Performance Factory, in which Codex, the company's coding AI agent, investigates each performance alert and proposes a fix. A second Codex session tries to disprove each diagnosis, and an engineer decides whether the fix is merged. In a manual review of 100 diagnoses, the team graded 70 percent as accurate.

Megan Chan and Rachel Banawa of Nielsen Norman Group (21 Nov · 02), a user-experience research and training firm, tested what an AI model should read before it designs. The AI model received either about 100 raw user comments or findings and recommendations that a researcher drew from the same comments. Designs built from the findings met 77.9 percent of a fixed scorecard, against 69.5 percent for designs built from raw comments. For a flight-rebooking form, no raw-comment design said whether loyalty miles were affected, and every findings-based design did. The authors conclude that research now has two readers: the team and its AI tools.

Luke Wroblewski (21 Nov · 03), a product designer, published his notes from Crazy Agent Scale, an event at the venture firm Sutter Hill Ventures. David Cramer, co-founder of the error-monitoring company Sentry, described Junior, an AI agent that does nearly every job at Sentry and makes the most commits. Junior opens pull requests under its own name and credits the person who asked on every commit. Quinn Slack, chief executive of Amp, a company that makes a coding AI agent, said the same 21 people took Amp past \$100 million in annual recurring revenue. At Amp, one person with an AI agent decides, ships, supports and markets each area, and a person still reads every customer email the agent drafts.

What they add up to

In all three accounts, people no longer write most of the work; they choose what the AI agent reads and approve what leaves the company. OpenAI replays past production incidents whenever it changes the agent's instructions, and Nielsen Norman Group scored every design against a written checklist. Both are ways to learn that an agent's inputs are wrong before an incident shows it, which is this issue's inference rather than either author's claim.

The case against

OpenAI is describing its own product, Codex, and the Nielsen Norman Group study used 32 designs on two familiar tasks. Teams with messier data and less tooling may see much smaller gains.

Receipts
OpenAI
How OpenAI’s Performance Factory keeps ChatGPT fastCompany story · 21 Nov 2026
Okay Zed and Cyrus Forbes · 19 November 2026

OpenAI engineers built a system in which Codex AI agents investigate and fix ChatGPT slowdowns, a second agent challenges each diagnosis, and an engineer decides what is merged.

Nielsen Norman Group
Can AI Generate Good Designs from Raw User Data?How-to · 21 Nov 2026
Megan Chan and Rachel Banawa · 20 November 2026

Nielsen Norman Group found that an AI model designed better interfaces from researchers' findings and recommendations than from raw user comments.

LukeW
Crazy Agent Scale: Companies All-in on AICompany story · 21 Nov 2026
Luke Wroblewski · 20 November 2026

Luke Wroblewski's notes from an event on large-scale AI agent use show Sentry and Amp running most of their work through AI agents, with people reviewing what goes out.