Working Surfaceissue · 8 Nov
Issue · 4 links
Issue

Sunday
8 November 2026

Four accounts describe AI agents that hand people only the decisions they cannot settle, with the facts and evidence attached. People review those decisions and that evidence instead of reading everything the agent wrote.

Swipe for 02 to 04
8 Nov
The note

Teams that use AI agents for most of their code are changing what a person reads before work ships. Instead of reading everything an agent produced, people settle the decisions the agent could not make and check the evidence it brings.

Ankit Jain (8 Nov · 01) is co-founder and chief executive of Aviator, which sells tools for shipping AI-written code. In a sponsored post on The New Stack, Jain described how the company's own team changed code review. After AI agents began writing most of the code, the nine engineers went from about 100 merged pull requests a week to more than 400. An AI reviewer, a four-hour review deadline and a tuned self-review tool all failed, because each still asked people to read the code line by line. The team turned a year of review comments into 64 rules that machines check on every change, and had agents prove each acceptance criterion with screenshots, requests and logs. Each pull request now lists every decision, where it came from and who settled it, and the company reports four times the merged work and 40 per cent fewer reverts.

Shimochi (8 Nov · 02), an engineer at Rescue Now, a Japanese company that supplies crisis-management information for disaster preparedness, wrote in Japanese about handing task management to a coordinating AI agent. The coordinator compares the engineer's GitHub and Notion work, passes ready tasks to worker agents and decides their order. It stops and asks before design choices, production changes and anything another person would see, such as a pull request sent for review. Each question appears in a side panel with the decision on the first line, the facts below it and then the options. The engineer answers with an item number and a word, and reports no longer searching for the next task.

Reito Yamaura (8 Nov · 03), an engineer at Offgrid, a Japanese company that mainly places engineers on client projects, wrote in Japanese about a month of building business software with Claude Code. Giving the requirements straight to the agent in the code repository mixed reading the specification with writing code, and the session filled up quickly. Yamaura split the work: one session reads the requirement documents and writes an implementation prompt, and a person reviews that prompt before a second session builds from it. The prompt forbids the agent to guess unclear business rules, so it marks each one in the code and lists it in its completion report. That report became the list of questions to take to the client.

Ben Holmes (8 Nov · 04), an engineer at Warp, which sells software for running teams of AI coding agents, set out three ways to cut the time people spend reviewing pull requests from agents. First, a review agent argues with the coding agent in rounds before any person looks, and hands a question back to a person only when a change goes beyond decisions already written down. Second, the agent tests its own work as a user would, on a virtual computer, and attaches screenshots and short recordings of the state before and after. Third, each pull request is scored for risk on five measures: how well it was tested, whether it stays within the task, how much it touches, how hard it is to undo and whether its errors would be seen. Holmes says none of this removes the human reviewer; it tells the reviewer where to look.

What they add up to

In all four accounts, the agent's work arrives with its open decisions set apart from the work it could settle alone. People spend their time on those decisions and on evidence, not on reading every line. The inference, which is not any author's claim, is that a team's main design question is which decisions an agent must hand back, and in what form.

The case against

Two of the four accounts describe one engineer's own setup, and the other two are vendors' accounts: Aviator's sponsored post about its own team and Warp's guide to its own product. None shows the method holding up in a larger organisation or under an independent check.<br />

Receipts
The New Stack
How we review 4x more AI code without reading every lineCompany story · 8 Nov 2026
Ankit Jain · 7 November 2026

Aviator's nine engineers stopped reading AI-written code line by line and now review intent, evidence and open decisions, merging four times as many pull requests.

Rescue Now Tech Blog
Leaving task management to AI agents too, until the only work left was answering their confirmation requestsHow-to · 8 Nov 2026
Shimochi · 6 November 2026

A Rescue Now engineer lets a coordinating AI agent run their task list and now answers only the decisions it queues, each set out as decision, facts and options.

Qiita
How development started to flow after separating the Claude that writes prompts from the Claude that implementsHow-to · 8 Nov 2026
Reito Yamaura · 6 November 2026

An Offgrid engineer splits AI work into one session that writes the build instructions and another that builds, and forbids the agent to guess unclear business rules.

Warp
Addressing the code review bottleneckHow-to · 8 Nov 2026
Ben Holmes · 5 November 2026

A Warp engineer has AI agents review and test their own code changes and score each pull request's risk, so people know which parts need their attention.