AI agents and AI tools now produce the stories, specifications and applications that used to come from product and engineering staff. The five pieces today ask who checks that output, who looks after it later, and how much an AI agent knows about the company around it.
Satya Nadella, the chief executive of Microsoft, told Gergely Orosz of The Pragmatic Engineer (17 Nov · 01) that code is now cheap to write and hard to verify. He described the Azure networking team, which manages more than 500 fibre network operators, now handling its email and processes with AI agents. He advised companies to treat AI agents as insider risks and to build a separate system that checks they work as expected. Microsoft has also offered voluntary exit to most of its software engineers, and Nadella said junior hiring has slowed while teams settle their new size and roles. His point is that the work of engineering is moving from writing software to checking it.
Ryan Narragon of Atlassian (17 Nov · 02) wrote up how Lendi Group, an Australian home-loan company, and CHG Healthcare, a healthcare staffing firm, use AI agents. At Lendi, Matt Hargraves, head of product delivery and automation, set one rule: all work lives in Jira, Atlassian's work-tracking product. Recorded video walkthroughs feed an AI agent that drafts the project document, and other agents break it into user stories with acceptance criteria written for coding agents. CHG spent two years on rules and oversight before putting any agent into production, and keeps its agent instructions as versioned files that can be rolled back. This is a vendor's account of its own customers.
Vishnu Elangovan, Anupa Bhattacharyya and Debaprasun Chakraborty of Amazon Web Services (17 Nov · 03) described six AI agents in AWS Quick, Amazon's workplace AI product, that handle the product work before coding. The agents gather customer complaints from tools such as Slack and Jira, score them, draft a press release and FAQ in Amazon's style, and write a specification with acceptance criteria. Each stage stops until a person approves it: the team confirms the problems, approves the priorities and signs off the specification. The authors argue that as coding speeds up, the decisions made before it matter more. The post reports no measured results and describes the company's own product.
Chris Stokel-Walker (17 Nov · 04), a freelance journalist, wrote in LeadDev, a publication for engineering leaders, about employees outside engineering who build their own software with AI tools. Retool, a maker of software for building internal tools, surveyed 817 people who build software. Of those, 60% had made something outside IT oversight in the previous year. Christina Troy, senior IT manager at the automation company Tines, said IT teams often learn of such software only when something breaks. Tines built a platform where employees build software while IT and security teams see which applications reach which tools and data. The article's point is that engineering teams inherit tools they did not design, so oversight has to start when the building starts.
Joshua Qin and Phil Chen of Concept (17 Nov · 05), an applied research lab, tested how well AI agents understand the company they work in. The two authors wrote 205 questions from one month of the lab's own messages in Slack, the workplace chat product. Each question asks who wrote a message, who was tagged in it, or who wrote each reply. The best of four AI agent products answered 75.9% correctly. The agents recognised who owns a responsibility far more often than what a person cares about. The authors argue that an agent can complete tasks well and still lack the knowledge of colleagues that an employee has. The test covers about 10 people at one company for one month.
In two of these accounts, AI agents write the stories and acceptance criteria that coding agents then follow. Lendi and Amazon place a person at the point where that output is accepted. Nadella argues for a separate check on whatever agents produce, and Troy for a view of whatever staff build. Concept's test gives one reason the check needs a person: the agents were weakest at knowing what colleagues care about. The inference, which no author states, is that a team should name the person who reads agent-written stories against the original intent before any code is built.
Two of the five accounts are vendors describing their own products, and a third quotes a vendor about its own platform. None reports how often a person rejected or rewrote a story the agents produced. Approval at every stage may become a formality once the drafts look good enough. Concept's test draws on about 10 people at one company, so its scores may not hold elsewhere.
Satya Nadella, Microsoft's chief executive, says code is now cheap to write and hard to verify, so companies should build separate systems that check their AI agents.
Lendi Group has AI agents turn project documents into user stories for coding agents, while CHG Healthcare spent two years on oversight before any agent reached production.
Three Amazon Web Services authors describe six AI agents that turn scattered customer feedback into a signed-off specification, with a person approving every stage before coding begins.
Staff outside engineering now build their own software with AI, a LeadDev report finds, and engineering teams inherit tools they did not design.
Concept, a research lab, tested how well four AI agent products understand a company from its chat history, and the best answered 75.9% of 205 questions correctly.




