AI Agents Shift Engineers From Coding to Governing Code
Microsoft’s Charles Lamanna argues that AI agents will matter less as add-ons to existing workflows than as a reason to redesign how work is organized. He says Microsoft engineers are moving from writing code line by line to directing and evaluating agent-generated output, while sales, finance and HR teams can increasingly produce their own analysis rather than wait for central reporting groups. Lamanna’s case is that the resulting productivity is more likely to be reinvested in growth and new work than used simply to reduce headcount.
AI’s advantage, Lamanna says, comes from redesigning work—not bolting on agents
Microsoft’s Charles Lamanna argues that AI agents change an organization most when they alter how work is organized, who can produce analysis, and where human judgment is applied. The alternative—adding agents to existing processes without changing those processes—will leave companies behind, he wrote in an August 5 blog post highlighted by Bloomberg.
The teams that redesign how they operate around AI will outperform those that bolt on agents to old ways of working.
The operating-model shift Lamanna describes has two related forms. In software engineering, agents generate more of the code while engineers spend more time directing, testing, coordinating, and approving the output. In business functions, tools that once required a central reporting team can move closer to the sales, finance, and HR leaders who need the analysis to make decisions.
Neither example is presented as a claim that human expertise becomes unnecessary. Lamanna’s argument is that the constraint moves: away from manually producing each line of code or waiting for a dashboard backlog, and toward defining what good output looks like, evaluating it, and deciding what to do with it.
Engineering shifts from typing code to governing generated code
? charles-lamanna describes the prior software-development model as one in which engineers wrote code line by line, making code production the central constraint. With agentic coding, he says, Microsoft engineers are increasingly responsible for overseeing systems that generate code.
Our software engineers are spending most of their time overseeing these agentic systems which will generate the code.
That oversight includes building verifiers and evaluations to measure quality, making architectural choices, coordinating work, and exercising “editorial” judgment over what should enter a product. Generating more code expands what a team can attempt, but it also raises the importance of those controls.
Lamanna expects the profession to look “nothing like it has in the past” by the end of 2026. Still, he does not characterize agentic coding as useful only for new projects. New capabilities benefit most, he says, but mature products can also see substantial efficiency and productivity gains.
His example was the on-premise SharePoint server, a 25-year-old codebase. Lamanna said someone on his team had described an “agentic coding revolution” there, with progress accelerating dramatically. The benefit, in his account, sits on a spectrum: agents can be especially effective in greenfield work without being confined to it.
Dashboarding moves from a central queue to the people making decisions
The same redesign appears in Microsoft’s use of data analysis, dashboards, and reports. Lamanna says the company had relied on central teams to build visualizations for sales, finance, and HR leaders. Copilot is beginning to let people in those functions conduct the analysis, assemble dashboards, and share the results themselves.
That removes a bottleneck around both production capacity and understanding. Rather than formulate a request and wait for a specialist team to work through its backlog, a finance, sales, or HR leader can create a report in an afternoon using Copilot, according to Lamanna. The person closest to the decision can also determine what information is useful to see.
Lamanna identifies two effects: less cost and investment devoted to dedicated dashboard production, because that work is distributed across the organization; and faster, potentially better decisions, because more people can develop insights without waiting for a central team.
Ed Ludlow put the staffing implication directly: if business users can perform analysis previously handled by specialized teams, might fewer people be needed in those organizations? Lamanna’s answer was that the pattern Microsoft is seeing is not necessarily team reduction. Companies, he says, tend to use surplus time and productivity to pursue growth.
Big thing that we’re seeing is not reducing the team necessarily, it’s really about doing more and interesting work.
That reinvestment can include researching new products, improving customer engagement and retention, or improving employees’ experience so they are more likely to stay. Lamanna’s premise is that businesses will often trade cost savings for top-line growth: agents allow an organization to undertake more work with its existing people rather than serving chiefly as a headcount-reduction mechanism.
Copilot’s growth depends on adoption inside existing work
Microsoft 365 Copilot had passed 30 million paid seats in the July period, Ludlow said, up from 20 million in the March quarter. Lamanna called the preceding six months a turning point for the product: Microsoft had taken a couple of years to reach 15 million seats, disclosed two quarters earlier, and then rose above 30 million.
Lamanna attributes the acceleration to successful initial deployments, rising engagement, model choice, workflow integration, and access to organizational context. He says the number of Copilot conversations more than doubled year over year, and that Copilot engagement is on par with Outlook and Teams—applications knowledge workers use regularly. Microsoft planned to track those measures through the rest of the year.
The first adoption driver he identifies is a multi-model strategy. Microsoft initially offered GPT and OpenAI models as the main choices customers wanted, then added Anthropic’s Claude, announced a partnership with Mistral in Europe, and began working with Microsoft AI and open-source models. Customers, Lamanna says, do not want to be locked into one model; Copilot offers a single application through which they can choose among models.
The second driver is placement. Copilot is integrated into Excel, Outlook, and PowerPoint, so users can work inside familiar applications rather than move to a separate destination or reconstruct their workflows around a new tool.
The final distinction, Lamanna says, is context. Every company can access the same underlying models, so model access alone does not differentiate one organization from another. Microsoft’s proposition is to bring together information in a customer’s mailboxes, files, documents, dashboards, and reports, then make that information available through Copilot. The useful combination, in his framing, is the model plus the context that is particular to a company.