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.
The $172k translator
Tech is still shedding jobs, but one title grew 729% in April: the forward-deployed engineer, averaging $171,911 a year. The role Palantir invented – an engineer who embeds with the customer, translates between the product and the business, and ships fixes from the client’s conference room – is now being copied by Amazon, OpenAI, Microsoft, and Salesforce. Kari McMahon’s reporting suggests the copies may be half-measures. Palantir’s own commercial head says every replicant he’s encountered is exactly that, and an ex-Palantir consultant admits the title “might have become a little bit frothy.” The demand signal underneath is real, though: what companies are paying for is technical-business bilingualism, the ability to sit with a customer’s mess and translate it into working software. That’s a skill, not a job posting, and it comes with a 50% travel load and a burnout profile that makes the salary look less generous. If you lead a team, note what just got publicly repriced: engineers who can actually talk to customers.
— LeadDev, 5m, #engineering, #careers, #strategy
Faster, not richer
Four years after ChatGPT launched, the company-level and economy-wide productivity numbers still show nothing – an awkward fact for anyone budgeting for AI next year. Matt Scherer’s explanation: we’ve been measuring the wrong thing. AI delivers what he calls casual productivity – tasks completed faster – while economic productivity is about the value of outputs relative to inputs, and value depends on quality. The Industrial Revolution’s machines made textiles faster and better. Generative AI makes work faster and then hands you the cleanup: fabricated case law, deleted databases, wrong diagnoses, workslop flooding inboxes. AI has doubled monthly app store releases while the number of apps people actually use has declined. More output, less value. Meanwhile $1.5 trillion in infrastructure spend needs to be earned back before anyone breaks even. You don’t have to buy his full pessimism to take the operating lesson: measure your AI initiatives by the value of what ships, not by how fast the tasks fly by. Lines of code are just that.
— Hard Reset, 10m, #ai, #productivity, #strategy
Nobody understands Google search
Not one person alive fully understands the codebase behind Google search, and it works anyway. Sean Goedecke uses that fact to defend something engineers feel increasingly guilty about: working in systems they only partly understand. The classic objection comes from Peter Naur’s “theory-building” – real understanding lives in programmers’ heads, so a team that loses the theory should rebuild from scratch. Fine advice when GCC was 100,000 lines. It’s 14 million now. Goedecke’s alternative is to understand one flow end to end, then expand outward as the work demands. Complete understanding, he argues, is one value among many, and it trades off directly against speed, legal compliance, security patches, and the ability to move people between teams. LLMs cut both ways here: they erode the deep mental model while making partial understanding dramatically faster to acquire. So stop treating “nobody fully understands this system” as an emergency. It’s the permanent condition of large software, and that’s not so bad.
— Sean Goedecke, 5m, #engineering, #systems, #ai
Coding agents, off the cloud
Sebastian Raschka wired up a fully local coding agent – open-weight model, no API bill, no data leaving the building – and documented every step. The stack is simple: Ollama serves Qwen3.6 35B-A3B on a Mac Mini M4 at about 40 tokens per second, and any harness that speaks the OpenAI API – Qwen-Code, Codex, even Claude Code – can drive it. The model solved four of five benchmark tasks, and harness choice mattered less than expected. Two findings deserve your attention even if you never go local. First, the security audit: Raschka inspects each harness for data egress and telemetry before installing anything, and finds Qwen-Code ships with Alibaba telemetry on by default. How many teams audit their cloud agents that closely? Second, the economics: local means fixed costs and reproducible behavior, which starts to matter when your monthly token bill has a comma in it. Frontier models are still better, and he says so plainly. But the floor is rising fast, and for regulated teams or anyone whose code can’t leave the premises, good enough just became a real option.
— Sebastian Raschka, 18m, #ai, #engineering, #tools
What the resistance is protecting
Your team isn’t resisting the reorg because they’re stubborn. Andi Roberts pulls together the behavioral research on why change stalls, and the through-line is that resistance is information. Sometimes it’s flagging a real flaw in the plan. And sometimes it’s a hidden commitment – Kegan and Lahey’s research shows people can sincerely support a change while unconsciously protecting themselves from what it costs them. Often it’s plain arithmetic: employees now absorb about ten planned organizational changes a year, up from two in 2016. The fifth reorg lands differently than the first. Roberts’ advice? Don’t sell harder – perceived fairness and honest information reduce resistance more reliably than enthusiasm or incentives. Don’t treat resisters as one bloc; six distinct profiles show up in the research, and your effort belongs on the movable middle, not the deniers. And before launching anything, ask what the change requires people to give up, because that answer predicts the resistance you’ll face.
— Andi Roberts, 10m, #leadership, #management, #transformation
Turn the lights back on
The “lights-off software factory” – agents writing, reviewing, and merging code with no humans in the loop – keeps failing, and Dex from HumanLayer has a structural explanation for why. Models are trained on feedback that arrives in seconds: tests pass or they don’t. Bad architecture sends its bill weeks or months later, far too slow to ever reach the training loop. So there’s no penalty for eroding maintainability, and the erosion is measurable: PR review quality falling since January, incidents per pull request up 242.7%, bugs per developer up 54%. His fix is unfashionable relative to the current hype train: keep human review, and make it cheap by front-loading agreement – product review, then system architecture, then program design, then vertical slices that can each be verified. Get the humans aligned on mockups, contracts, and pseudocode before the agents generate a line of code. He’s candid about his incentives (his company sells human-in-the-loop tooling for devs), but the argument stands on the constraint itself: “embrace the constraints and move 2-3x faster, safely,” instead of chasing 10-100x into a rewrite.
— HumanLayer, 25m, #ai, #engineering, #systems
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.
Captain’s log turns your team’s week into a formal ship’s log entry: goals become your heading, accomplishments the distance traveled, blockers the weather encountered, and check-in morale the condition of the crew. Every Friday at 5 PM it enters the week’s voyage into the record – “heavy seas near the payment integration; crew morale steady despite the loss of two days to the staging outage.” It reads like fun, and it is. But it quietly builds the archive every team wishes it had: months later, when someone asks why you changed course, the log remembers.
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.