Working Surfaceissue · 18 Nov
Issue · 5 links
Issue

Wednesday
18 November 2026

Five accounts from 14 to 17 November show teams whose AI agents and AI tools now produce steady work turning their attention to where that work stalls or must be checked. An engineering leader, an Intuit engineering manager, a Salesforce team, PostHog's designers and Figma's TJ Pitre each describe where people now stand in that flow.

Swipe for 02 to 05
18 Nov
The note

As AI agents raise how much code a team produces, the hard work moves to seeing where work stalls and checking what the agents made. Five accounts published between 14 and 17 November show engineering, product and design teams placing people at exactly those points.

Cate Huston (18 Nov · 01) is an engineering leader whose small software team has AI agents built on Anthropic's Claude write most of its code. Output held at about two pull requests an hour, yet work stalled when nearly finished and nobody could see what was blocked on whom. Huston cleared the queue of ready work and gave each large feature a written specification, a feature flag, a rollout checklist and its own Claude agent to hold its context. Huston reports about 50 pull requests a day, and says November is on track for roughly double the output of September and October.

Arjun Garg (18 Nov · 02), a senior engineering manager at the financial software company Intuit, asked why delivery stayed flat after AI tools were adopted. Garg compared spending on AI credits with merged pull requests and ticket closure. The heaviest spenders spent about twice as much as a middle group, but did not merge more pull requests. Garg proposes a test: give an AI agent a hard but bounded task, with a person setting the design and review standard, and see whether it stalls.

Shireen Nagdive (18 Nov · 03), an engineering manager at Salesforce, leads a team retiring an old login method for the SOAP API, an interface other software uses to connect to Salesforce. Enforcing the change broke more than 600 tests across several teams, and real faults looked the same as expected breaks. With Riya Suri, the team built a tool with Claude Code in which ordinary code fetches and routes failures, the AI model reads unfamiliar errors, and engineers approve each conclusion. Triage that was estimated at several engineers for two to three days took one engineer about 90 minutes.

Cleo Lant (18 Nov · 04) of PostHog, which makes product analytics software, explained a redesign of its web app. Nearly half of what customers create or change in PostHog now happens through AI agents. These agents connect through the Model Context Protocol (MCP), a standard way for agents to use other software. Yet about 40,000 of those customers returned to the web app within an hour of an agent's work, mostly to check dashboards and session recordings. The team replaced a list of pages with a short daily briefing, and cut the sidebar to five items built around questions users ask.

TJ Pitre (18 Nov · 05), Figma's product architect for AI and design systems, wrote in his newsletter Slot Machine about designers who now hand engineers working code made with AI. A component that looks and behaves right in a test playground can arrive with hard-coded icons and code that ignores the company's standards, and the engineer reviewing the pull request has to fix it. Pitre's answer is to hold designers' code to the same bar as engineers' code and run automated checks before any engineer opens the pull request. He would also give designers a lighter setup with a design system the AI can read. He argues that checks cannot judge whether the direction is right, which still takes a person and a design critique.

What they add up to

In each account, AI agents or AI tools produce more work than before, and the team's attention moves to the places where that work stalls or must be checked. Huston watches the flow, Garg tests whether the tools explain flat delivery, Salesforce engineers approve the tool's conclusions, PostHog designs a screen for people checking agents' work, and Pitre asks for checks that run before designers' code reaches an engineer.

The case against

Each account comes from one team or one writer, three are told by the people who built the system, and none shows that quality held as output rose. Garg's own data suggests that more AI use does not by itself bring more delivery, and Pitre draws on conversations rather than measured results.

Receipts
Accidentally in Code
Seeing the bottlenecks in the software factoryCompany story · 18 Nov 2026
Cate Huston · 17 November 2026

Cate Huston, an engineering leader whose AI agents write most of the team's code, now spends the working day finding where work stalls before release.

LeadDev
Stop blaming AI tools for flat velocityCompany story · 18 Nov 2026
Arjun Garg · 17 November 2026

Arjun Garg, a senior engineering manager at Intuit, found that the heaviest AI spenders did not merge more pull requests, and proposes a test to learn why.

Salesforce Engineering
How AI Slashed 600-Test-Failure Triage From Days to 90 MinutesCompany story · 18 Nov 2026
Shireen Nagdive and Riya Suri · 17 November 2026

A Salesforce engineering team cut triage of over 600 failing tests from days to 90 minutes, with code doing routine steps and engineers approving the AI's conclusions.

PostHog
We accidentally launched our new UIHow-to · 18 Nov 2026
Cleo Lant · 17 November 2026

PostHog redesigned its analytics app around a short daily briefing after finding that customers who work through AI agents still return to the app to check results.

Slot Machine
Design Engineering Has a Handoff ProblemHow-to · 18 Nov 2026
TJ Pitre · 14 November 2026

TJ Pitre, Figma's product architect for AI and design systems, finds that code designers make with AI now lands on engineers to fix, and says checks must run before review.