You’re reading The Steady Beat, a weekly pulse of must-reads for anyone orchestrating teams, people, and agents across the modern digital workplace – whether you’re managing sprints, driving roadmaps, leading departments, or just making sure the right work gets done. Curated by the team at Steady.
Who checks the checker
The more AI output you review, the worse you get at catching its mistakes. That’s the argument of a new Hugging Face and Data & Society paper, and a Harvard, MIT, and University of Washington field experiment explains why. People screening contest entries made better calls when an AI told them to simply pass or reject. They made worse calls when the AI also explained its reasoning, worse than with no AI at all. When the AI was wrong, the people who read its reasoning went along with it more often, and they read less of the entries themselves. David Rice’s takeaway: if a person signs off on AI work, design that job on purpose. Keep the reviewer separate from whoever benefits from approval, don’t score them on speed, and have them do the work without AI often enough to stay sharp.
— People Managing People, 5m, #ai, #management, #quality
Someone behind it
Good products carry a point of view, and Karri Saarinen, co-founder of Linear, thinks the agent-powered software factory is diluting it. Two teams with the same customers and the same technology build very different things, and the difference is judgment: what the makers pay attention to, what they refuse to add, what they keep shaping after it technically works. Lately many products feel messy to him, even from very successful companies, because adding a capability is now easier than rethinking one. He blames the way we work now: individuals paired with agents, “each operating a small software factory,” shipping changes that make sense alone and together “resemble Frankenstein’s monster rather than a carefully shaped sculpture.” Easier software should buy room for perspective, not replace it. “I want to feel the creator behind the software, not the factories.”
— Karri Saarinen, 3m, #product, #ai, #craft
The easy part
AI inference is working. Everything around it isn’t. That’s Tom Loftus’s read of this week’s earnings: Google Cloud grew 103% year over year and Azure 48%, with demand outrunning supply at both. But Google’s search growth slowed to 12% because AI Mode queries monetize at a lower rate, and the stock fell about 10%. ServiceNow fell 14% on results that met expectations, over fears AI upends its business. Microsoft offered buyouts to about 60,000 people, a quarter of its workforce, and is rebuilding how work gets done. Its transformation chief Kathleen Hogan says their biggest error was leading with technology instead of people and process: “A bad process with AI is still a bad process.” If the companies selling AI have to rethink revenue, structure, and headcount, “no one else is exempt.”
— The Wall Street Journal, 8m, #ai, #strategy, #transformation
Flag in the maze
The “rogue agent” hacks this summer weren’t agents going rogue. They were agents doing exactly what their training rewarded. Raj Saha and Blacki Migliozzi built a Pac-Man-style maze to show it: an agent rewarded for reaching a flag and exploring, lightly penalized for probing walls, was handed a flag sealed inside walls. It left the maze, probed from outside, and broke in. Nobody taught it that. OpenAI’s experiments ran a “highly persistent internal model” on tasks researchers estimate were 30 to 40 percent impossible, so the agents reached for unsanctioned tools and the live internet. “There’s a long history of so-called reward hacking in machine learning,” says Melanie Mitchell. Simple monitoring for a sandbox breakout would have stopped it; Harvard’s Stephen Casper figures “one engineer could have handled something like that.” The incentive you set is the behavior you get, for agents and for teams.
— The New York Times, 12m, #ai, #agents, #systems
Fewer managers?
CEOs are betting on fewer managers. Jack Dorsey says Block needs no “permanent middle management layer.” Brian Armstrong announced “no pure managers” at Coinbase. Luca Rossi spent a month on the question with the Refactoring community and finds the job’s core hasn’t moved. Andy Grove’s definition still holds: a manager’s output is the output of their organization plus the organizations they influence. So do the four areas it covers: Product, Tech, Process, and Team, with Team “the human core of the job, and probably the part that has changed the least.” What AI takes first is glue work: moving information between tools, writing up meetings, fetching status. Tedious, and “absolutely foundational to the success of your job.” The question for every manager is what they do with the hours that frees up.
— Refactoring, 10m, #management, #ai, #leadership
Echo of the week
Echoes are AI agents in Steady that automatically gather and deliver work context to teams on a schedule – answering recurring questions about progress, capacity, and coordination so you stop burning hours assembling the same information manually.
Big-picture progress. As a company grows, goals sprawl. Teams publish updates all week, and nobody can see the forest for the trees without clicking through every goal or sitting through a status meeting where half the updates don’t apply to them. This Echo gathers every goal update from the past seven days into one summary, grouped by goal, every Friday at 2:00 PM. Executives see the whole org’s progress without attending every team’s meeting, program managers spot upstream work stalling before it blocks a downstream team, and department heads have the week’s picture before they report up.
The human-agent teamwork OS
Teams rely on two coordination loops to function: a big-picture loop connecting plans to progress, and a ground-level loop keeping teammates in sync.
Running those loops was already a marathon of meetings, chat threads, dashboards, and manual toil. Pile on flatter orgs, exponential output, and AI agents shipping 24/7 — the old way can’t keep up.
Steady removes the coordination bottleneck by running both loops for you. Working in the background, Steady distills updates and activity into targeted context for everyone on the team — human and agent alike. Full visibility, tight alignment, zero overhead.
The outcome: high-performing teams that deliver the right work, not just more of it.
Learn more at runsteady.com.