Software development job postings are a quarter below pre-pandemic levels, yet active IT job postings hit a three-year high; tech layoffs are up 50%, yet Accenture's headcount reached 814,000. These numbers are all true simultaneously. Unpack IT work patterns—an engineer's day, roles and teams, the entry ladder, enterprise IT departments, IT services sold by person-days—and the confirmed changes are smaller and stranger than the narrative.
IT work patterns are indeed changing, but the shape of change is three shifts, not "people being replaced": work moves from writing code to orchestration, review, and writing specs—and it densifies, it doesn't shrink; the entry ladder loses a rung; the pressure is at the entry level, with no clear AI impact visible in aggregate numbers; enterprise IT gains a new function, building identity, permissions, and budgets for agents. Most evidence supporting these judgments is self-reported,0 vendor-sourced, or correlational.
Adoption rate is not output; job postings are not employment; company statements are not causal attribution. ByteDance's TRAE team has 90% of code written by AI, yet per-capita requirement throughput only increased ~60%; that gap is closer to the truth than any conclusion that "IT is disappearing" or "IT is fine."
In mid-August 2026, Indeed's software development job posting index was 74.4 (Feb 2020 = 100), still about a quarter below pre-pandemic. Almost simultaneously, CompTIA counted (Sept 2026 data) 626K active IT job postings, +22.4% YoY, the highest in over three years. Tech sector layoffs announced Jan–Aug 2026 were 155,126, +52% YoY. Accenture announced up to $865M in restructuring in Sept 2025; one year later, its earnings press release reads "approximately 814,000 people."
None of the four numbers is wrong, and no two of them combine into "the IT industry is shrinking" or "recovering." Job postings measure demand intent, layoff announcements are companies' stated plans, headcount is the employed stock, and the base periods are all different.
So "how IT work patterns change in the AI Native era" cannot be answered with a single number. The previous article, "R&D in the AI Native Era," argued that after writing becomes cheap, the scarce resource is the verifier. This article asks from the other end: after writing becomes cheap, what are people in IT doing, how do they divide labor, how do they organize, how do they enter the field?
I break it into five pieces, then look at one more thing separately—how managers get people to use AI. Each piece separately covers three types of content: what's confirmed, what's disputed, and what's merely narrative.
Start with an engineer's day. Anthropic used its own engineers as a sample and provided a rare set of "task structure" data: in Claude Code usage logs, "implement new feature" share rose from 14.3% in Feb 2025 to 36.9% in Aug, "code design or planning" from 1.0% to 9.9%. Engineers self-reported productivity +50%; 27% of Claude-assisted work was "things they wouldn't have done otherwise." In the same survey, over half said the work they could fully delegate to Claude was only 0%–20%.
These numbers depict not people leaving, but people repositioning: more new features, more design and planning, and the majority still requires human supervision.
Pragmatic Engineer asked 906 readers in early 2026: 75% use AI for at least half of engineering work, 55% regularly use AI agents. The series' first installment summarized the trend in one line: engineers need to orchestrate and switch contexts more frequently, and managers can be more hands-on; the two roles are converging. This is a one-sentence summary from the newsletter intro, not quantified; the sample is newsletter readers, predominantly senior engineers in Europe and America with 11–15 years median experience—cannot be generalized to the whole industry.
Where did the saved time go? Several independent data points point in the same direction: work densified, it didn't decrease.
These four studies cover different populations and different measurements; they cannot be summed. DX's "6 hours saved" is self-reported, the same class of problem as METR's "feels fast, measures slow" (discussed in the previous article). But they paint a consistent picture: saved time became more tasks and more checking.
The largest chunk of this extra work is review. Sonar asked 1,100+ developers in Jan 2026: 96% don't fully trust AI-generated code's functional correctness, but only 48% say they always check before committing; 38% find reviewing AI code more effortful than reviewing colleagues' code (Sonar sells code quality tools—vendor-sourced). Empirical evidence on agent-submitted PRs is more direct: most AI-generated PRs have no review at all, and for those that do, reviewers are primarily AI agents. This echoes the CMU and Stanford finding from the previous article: human review coverage dropped from 89% to 68%, review became the new work主体, and simultaneously the most understaffed环节.
So how common is "one person managing a swarm of agents" in practice? The answer depends on whose definition you use.
All four numbers are vendor-sourced, with different time points and denominators; the same phrase "agent involvement" means different things at each company. Anthropic's own measurement of Claude Code shows a median single-turn duration of only ~45 seconds; no company has publicly disclosed the distribution of "how many agents each developer manages concurrently." The "3 to 5 parallel sessions" that can be found is a recommendation from the Claude Code team and the creator's personal habit.
Is "managing agents" a new skill? The only controlled experiment comes from NBER: leaders managed both human teams and AI agent teams; performance correlated at 0.81, and ability to manage AI is highly correlated with and transferable from ability to manage people (the task was collaborative information integration, not coding). A BCG survey (1,488 US employees, via The Decoder) found ~14% reported "AI brain fry," with the strongest driver being agent supervision; interestingly, in the same survey, people whose routine tasks were taken over by AI saw burnout scores drop 15%. What's exhausting isn't using AI—it's watching AI.
Role boundaries are loosening; several independent线索 point in the same direction.
LinkedIn CPO Tomer Cohen announced at the end of 2025 the end of the long-running APM (Associate Product Manager) program; starting Jan 2026, new hires enter as "Associate Product Builder": teaching them to write code, do design, and do product, replacing functionally divided large groups with cross-functional small pods. This is one company renaming a program, not industry data, but it aligns with the HBR field study finding that "product managers start writing code."
SignalFire's 2026 talent report is unusual in putting organizational structure metrics in a data report: big-tech engineering managers supervise ~12 engineers (previously 10), startups ~15; PM hiring at big tech is −39% vs 2019, at startups +2%; within engineering, AI/ML engineer share is +39% since 2022, frontend engineer share −25%. Note these are "share of engineering hiring changes," not absolute numbers; data comes from a VC firm's talent database, not official statistics.
New titles are also appearing: ByteDance's 2027 campus recruiting added "AI Full-Stack Engineer" and "AI Agent Development"; Forward Deployed Engineer postings on Indeed grew from 643 to 5,330 (+729%). But this is a percentage on a small base; Gartner in Sept 2026 also predicted that by 2028, 70% of enterprises will abandon vendor-built agentic AI deployed by forward-deployed engineers, warning of "FDE washing" (repackaging traditional consulting as FDE). As for Evaluation Engineer, Agent Ops and similar titles, I found no first-hand posting volume statistics—only observable phenomena, not "grew X%." The FDE story itself is covered in "FDE Manual" on this site.
So my judgment for this piece: role boundaries loosening is real; the scale of "new roles" mostly lacks data. It looks more like职责 consolidation and share migration than a batch of new occupations appearing from nothing.
"AI makes teams smaller and flatter" is the most common narrative. The evidence needs to be separated: half is observation, half is prediction.
What's observed: Gallup measured US managers' average direct reports rising from 10.9 in 2024 to 12.1 in 2025 (median still ~6; growth mainly from 25+ person super-teams), but the Gallup report mentions AI only once, saying "AI's future impact on team size is uncertain." Google's VP of People Analytics said at an Aug 2025 all-hands that managers overseeing fewer than 3 reports decreased by 35%; the reporting attributed this to efficiency, not to AI.
Amazon is the clearest example, because the same company made three inconsistent statements:
The prediction half: Gartner predicted in Oct 2024 that by end of 2026, 20% of organizations would use AI to flatten structures, eliminating over half of existing middle-management roles. That window is nearly closed; no one has demonstrated "half of middle management replaced by AI."
Evidence for smaller teams has the same structure. Carta data (via PitchBook) shows Series D companies average 131 people, down 29% from the 2023 peak; but Carta's own analysis says hiring was bound to decline after the 2021 bull market, and AI merely accelerated it. AI-native companies' revenue per employee is more striking—Lovable hit $400M annualized revenue with 146 people (TechCrunch)—but this is company self-reporting, survivorship bias, and the same article notes they're still hiring for 70 roles.
DORA 2025's "seven team archetypes" is often cited to show team形态 differentiation; caveat: this is a pre-AI team health profile, and its conclusion is "AI is an amplifier"—strong teams get stronger, weak teams get weaker. You can use it to argue differentiation, but not形态 change.
Flattening is real; attributing it to AI requires passing the attribution test first.
Evidence pointing to "entry-level under pressure" comes from five different sources:
Counter-evidence is equally first-hand and equally hard:
These two categories of evidence don't contradict each other; they measure different slices: Yale and the NY Fed look at national or regional aggregates, SignalFire and Stanford look at the entry level; one says "no clear aggregate impact," the other says "entry level under pressure." Placed side by side, the more stable statement is: entry level under pressure, no clear aggregate impact, attribution unproven. Writing "AI eliminated junior roles" oversteps the evidence.
What the ladder lost isn't just hiring slots, but the function of the rungs themselves. In an Anthropic randomized controlled trial, 52 mostly junior software engineers learned a new Python library: the AI-assisted group scored 50% on comprehension tests, the manual group 67% (p = 0.01); the AI group was ~2 minutes faster, but not significantly. Usage pattern matters: those who purely delegated, developed stepwise dependency, or had AI repeatedly debug averaged under 40%; those who generated then追问, asked for code with explanations, or only asked concepts averaged 65%+ (each pattern had only 2–7 people; causality not proven). In Pragmatic Engineer's second article, directors reported two phenomena: junior engineers are often the biggest AI token spenders; senior engineers start using AI themselves for work previously assigned to juniors and contractors—junior practice opportunities are being taken away.
Enterprise IT departments—service desk, ops, security, budgeting—are not the same as internet companies' R&D teams, and they're where vendor narratives are loudest.
Service desk. After acquiring Moveworks in late 2025, ServiceNow claimed its internal AI agents can autonomously resolve 90% of IT requests. VentureBeat's Feb 2026 coverage itself noted the methodology gap: "autonomously resolve" is undefined, how 90% was measured is unspecified, and it only covers ServiceNow's own employee requests. I found no independent (non-vendor) benchmark for autonomous resolution rates on real IT tickets. The closest independent evidence comes from customer service: an NBER study of 5,179 agents using AI assistants found issues resolved per hour +14% on average, +34% for novices—AI assisting humans, not replacing them.
Gartner analyst Nate Suda offered a qualitative note in March 2026: the IT roles first affected are "workflow heavy," specifically service desk roles, business analysts, project managers; in the same piece, Gartner tracked 1.4M layoffs in 2025, fewer than 1% attributed to AI efficiency gains. Gartner's CIO survey also gives an expectation: by 2030, 75% of IT work done by human + AI, 25% by AI alone, while considering AI's employment impact through 2026 as "neutral." This is CIO expectation, not a measured result.
Ops. AI SRE agent product announcements almost universally use framing like "from hours to minutes." Independent benchmarks paint a different picture:
ITBench-AA's exact words: "All frontier models score below 50%"; in OpenRCA 2.0, of cases where the root-cause service was identified, only 61.5% could be traced to a verifiable causal path; Sept 2026's EviRCA pushed OpenRCA's scores above 40% through engineering, not larger models, but it's a preprint. All of these measure "localizing root cause," not "fixing," and certainly not "autonomous production execution": in several vendors' own descriptions, the default actions are investigate, suggest, and produce a PR; production changes still require human approval. This connects to two articles on this site about permission boundaries: "What to Audit, Whether to Allow" and "Who Reviews Before Release."
Governance. This is the most concrete new function in enterprise IT: managing agent identity, permissions, and lifecycle. The numbers are all vendor-sourced; mind the definitions. CyberArk's 2025 survey found machine identities outnumber human identities 82:1, but this is the ratio of all machine identities (API keys, service accounts, automated processes, plus AI agents), not just AI agents—VentureBeat's headline got this wrong. Gartner said in April 2026 that only 13% of organizations believe they have appropriate agent governance, and offered six steps: governance policy, centralized agent inventory, identity/permissions/lifecycle, AI information governance, monitoring/correction, and encouraging responsible use. Microsoft productized this: agents are an account type in Entra; extending security functions to agents requires Agent 365 licensing, included in Microsoft 365 E7. The governed objects expand from employees to software entities, simultaneously productized and priced. The people managing this were partly covered in "The Rise of the CAIO" on this site.
Budget. Gartner raised its global IT spending forecast to $6.37 trillion in July 2026, +14.2% YoY; the increase comes mainly from data center systems (+62.5%) and IaaS (+29.3%)—compute and hardware, not labor-intensive services. IT budgets rising and IT labor share declining can both be true: money is flowing to compute. This line is explored in more detail in "From Seats to Compute" on this site.
Companies selling services by person-days should be the first place where signals of "AI replacing IT people" appear. The signals are more complex than the narrative.
TCS announced in July 2025 it would reduce headcount by ~2% (~12,000 people) this fiscal year. The CEO's explanation was "skill mismatch" and "unable to deploy," not AI-driven 20% productivity gains. Looking at quarterly earnings headcount: June 2025 was 613,069, Dec 2025 was 582,163, a net reduction of ~30,900—over 2.5× the announced number (includes attrition not backfilled; different methodology); after that, three consecutive quarters of growth, reaching 598,056 in Sept 2026.
Accenture announced up to $865M in business optimization in Sept 2025; CEO Julie Sweet's framing was to "exit people" who can't be retrained on a tight talent timeline. One year later, the earnings press release reads "approximately 814,000 people," FY2026 revenue $74.18B, 5% constant-currency growth. On the earnings call, the CFO said fixed-price work (including outcome-based) was 65%+ of new bookings; note this is the "fixed-price" category overall, not the outcome-based share specifically.
Campus hiring hasn't stopped either. Cognizant's CEO said in June 2026 that they hired 20,000 entry-level fresh graduates in 2025 and will hire more in 2026, calling predictions of job collapse "a little bit of fearmongering," while also saying "the middle layers are going to be leaner" (Fortune); Infosys's CFO said they'll hire at least 20,000 fresh graduates in FY27. These are hiring targets, not actual onboarded headcount.
Where did "pyramid to diamond" come from? It's been a metaphor in analyst circles for a while; a Gartner research piece is literally titled "From Pyramids to Diamonds"; a Forrester analyst corrected in July 2026 that the diamond's base "looks more like an upright pentagon" (The Week). It's an analyst prediction and metaphor; no services firm has disclosed actual headcount by layer; Cognizant's CEO simultaneously said "wider pyramid" and "leaner middle layer," showing the shape depends on who you ask and in what context.
Evidence for a shift from person-days to outcome or token-based pricing is equally thin. HFS's survey of 25+ large providers found revenue/FTE up only ~1.7% YoY, operating profit per FTE up ~6.5%—"decoupling" exists but the magnitude is small and concentrated in a few companies; HFS itself is the main promoter of this narrative, profiting from selling related research and awards.
Two Chinese samples appear directionally similar, but mind the methodology. iSoftStone's 2025 annual report states that, in response to large models' impact on traditional "project-based" delivery, it adopted an "AI Factory + Industry Know-how + FDE" delivery model, with 91,849 employees at year-end. Its "AI-related business" accounted for 61.46% of H1 2026 revenue, sounding like AI transformed the services; but broken out, the driver was computing products and smart electronics (+40.32%, 7.70% gross margin, mostly hardware), while software and digital technology services grew only 0.56% with gross margin down 2.98 percentage points. Chinasoft International had 78,030 employees at end of 2025, 2,955 more than the prior year; per its July 2026 voluntary HKEX filing (via Sina Finance), it signed a token revenue-sharing cooperation agreement with Moonshot AI to jointly build an FDE innovation lab; commercial terms were not disclosed.
From 2025 to 2026, a batch of companies turned "use AI" into a management requirement. Several examples with direct quotes:
Pragmatic Engineer's second article noted companies setting 50%, 80%, 100% AI usage rate targets, treating such targets as danger signals. A roundup in domestic media also had anonymous employees saying they "waste tokens" to hit token quotas—anonymous hearsay, not any company's official policy.
Microsoft's 2026 Work Trend Index (20,000 AI-using knowledge workers, 10 markets) provides a contrast: only 13% said they were rewarded for using AI to reshape work, even when results weren't achieved; organizational factors explained about twice as much AI impact variance as individual factors (~67% vs 32%, statistical association, not causation). Microsoft sells Copilot—vendor-sourced. Putting it all together: requirements arrived; reward mechanisms did not keep up. The evidence shows a shift in management intent, not proof of effectiveness; the intent itself is oscillating—Duolingo reversed within a year.
Going through this round, I kept hitting the same four traps. I'm writing them here because they're more worth taking away than any single number.
Below are six practices I've distilled from this evidence. Some are directly supported by evidence; some are my judgment. I've marked which is which.