Writing / 2026
The Flattened-Management Trap
Removing a management layer leaves work behind. Assign it explicitly before hands-on managers inherit it alongside their coding commitments.
Amazon’s memo called for increasing the ratio of individual contributors to managers by at least 15% by the end of Q1 2025 . It complained about “pre-meetings for the pre-meetings for the decision meetings”. I support deleting that work. The harder part is deciding which responsibilities survive the person whose calendar held them.
Gartner forecast that through 2026, 20% of organizations would use AI to flatten their structures, eliminating more than half of current middle management positions . That is a forecast. Google’s reported 35% manager reduction needs a different qualification: it covered managers with fewer than three reports, and many stayed as individual contributors. Neither establishes that engineering organizations can remove management work wholesale.
My inference is that the surviving work slides toward hands-on line managers and tech lead managers, or TLMs, who are also expected to ship with agents. No study in this evidence measures that migration. It is the organizational risk I want leaders to test: a deleted position becomes an unstated addition to somebody else’s job, while their delivery commitment stays intact.
Will Larson, CTO at Imprint, captures the tension. He praises managers contributing software when they validate the work directly . In his interview , he calls the TLM role “a highly demanded trap” and judges line managers poorly on the bureaucratic dimensions. That last point is his judgment. His earlier writing describes a narrower mechanism among first-time managers : people doing both jobs often retreat toward familiar work and forget the management part. Shipping code with agents can be valuable while still giving a manager somewhere attractive to retreat.
In my engineering manager versus tech lead post , I wrote that a manager writing production code meant something was wrong. That rule has changed. I now support managers contributing code when they can validate it and make room for it. What survives from the old rule is the warning about substitution. A shipped change cannot stand in for a management responsibility that nobody performed. Agent assistance does not make those responsibilities optional.
Larson’s account of middle management gives that responsibility a useful shape. He describes balancing executive, stakeholder and team pressures; defining and operating organizational process; and competing for a share of fixed resources. His article concerns career paths. Applying those jobs to organizational design is my recommendation. Delete the layer if it helps, then decide where each job goes. “The teams will handle it” leaves the allocation decision unfinished.
Google’s 2002 experiment supplies a concrete failure. After removing managers, it relented when too many people went directly to Larry Page about expense reports and interpersonal conflicts, among other issues. Technical judgment alone also deserves proportion: the New York Times reported in 2011 that technical expertise ranked last among Google’s eight manager behaviors . That finding concerns that study. It gives no license to treat technical competence as irrelevant to managing engineers.
Part of the case for flattening holds. Much of the removed layer may have been overhead. Some coordination work is automatable. Gartner expects AI to automate tasks, reporting and performance monitoring , and Google deliberately operated with engineering managers carrying 30 direct reports . I favor moving decisions closer to the information . Restoring every approval would defeat the point. Keep the deletion where it removes work; assign whatever still requires judgment.
Gallup supplies a warning about capacity, with limits. Its U.S. research across industries reports that managers spend a median of 40% of their time on individual contributor work. Beyond that threshold, engagement is lower, and the relationship worsens as the number of reports increases. This is an association, outside engineering specifically. Gallup also reports manager engagement at 22% in 2025 . Those findings do not establish that flattening caused the decline. They do make unlimited coding expectations alongside wider spans a poor planning assumption.
Apply the ownership principle from my AI leadership bench essay to the work left behind. Give each cluster of TLMs someone with authority to resolve competing pressures. Assign organizational process explicitly. Put resource allocation with someone who can negotiate a budget and defend the decision. These jobs can share an owner. Choose that owner for the work, and check their capacity before adding it to their remit. Do not make strong coding performance the qualification for owning process or budgets. Ask the proposed owner to explain how they will handle a disputed allocation, and give them support where their experience falls short.
Budget the TLM’s hands-on time after those obligations have an owner. Keep Larson’s condition : managers contributing software must validate the work directly. Include that validation in the time budget. If the budget cannot accommodate it, reduce the coding commitment. Treat 40% as evidence to investigate workload, rather than a universal safe limit or an engineering target.
Next week, take the last quarter’s cross-team escalations and find where they landed after the layer disappeared. Follow the unresolved budget requests too. Use issue-to-decision time to check whether the new arrangement actually works, then move one stranded responsibility to a named owner with the authority and time to finish it.