Durable Skills

Confidence 0.95 · 43 sources · last confirmed 2026-08-30

The cluster of human skills that remain economically valuable as AI systems take on more codified, procedural, and information-retrieval work. The canonical examples — repeated across the educational, OECD, and labour-economics literatures — are collaboration, creativity, and critical thinking, sometimes joined by communication, problem-solving, self-regulation, and adaptability.

The wiki’s anchor source is Globerson et al. (Google Research, 2026), which uses the term durable skills in its title and operationalises measurement for three of them. The terminology is not unique to that paper — the same construct appears in adjacent literatures as 21st-century skills, future-ready skills, transferable skills, or simply human skills. This concept page treats them as substitutable for now and will note vocabulary drift if/when sources push back on the equivalence.

Working definition

A durable skill is a human skill that is:

1. Open-ended. Not specifiable as a closed-form rule or a deterministic procedure (otherwise an AI system can encode and execute it).
2. Socially or contextually situated. Performance depends on grounding in a specific group, organisation, or domain in ways an AI system cannot reproduce without that context.
3. Slow-changing. The skill itself doesn’t go stale every 18 months the way a specific tool or framework might.

The term explicitly contrasts with codified knowledge (factual content, procedural rules) which is far more substitutable by foundation models. The wiki’s ai-deskilling concept tracks the substitution direction; this page tracks the inverse — which skills humans should retain and how to measure them.

The measurement problem

Durable skills are notoriously hard to measure at scale. Three reasons recur in the literature (Globerson et al. §1):

1. Definitional fluidity. Different frameworks operationalise “collaboration” or “creativity” differently. The major prior efforts — PISA 2015 (scripted simulated teammates, multiple-choice answers) and ATC21S (Assessment and Teaching of 21st Century Skills, human-human dyads in a digital environment) — span only a small slice of the validity-vs-rigor tradeoff.
2. Group-task interdependence. Assessing an individual in a collaboration setting is psychometrically hard — other team members introduce uncontrolled variance.
3. Evidence scarcity. As Sijtsma (2011) put it, “long-lasting observation of a person in real life until (s)he spontaneously exhibits the behavior of interest would take too much time before enough evidence was collected.” Authentic interactions are evidence-thin per minute; structured assessments are evidence-rich per minute but artificial.

The historical compromise was highly scripted AI teammates (PISA) or highly structured human dyads (ATC21S). Both sacrifice authenticity to measurement reliability.

Vantage / Executive LLM — the 2026 measurement breakthrough

Globerson et al. introduce Vantage, an LLM-orchestrated assessment platform that resolves the validity-vs-rigor tradeoff via two architectural choices:

• Single Executive LLM driving multiple AI teammates. Rather than independent agents, one LLM generates responses for all AI teammates and is prompted to maximise information about the participant’s skill level — analogous to a computerized adaptive test (CAT) but for complex social/cognitive constructs rather than item difficulty.
• Separate AI Evaluator. A different LLM scores the conversation transcript against multi-dimensional rubrics (linear regression for scores; logistic regression for “no-evidence” labels), with leave-one-out cross-validation.

Validation (large-N): 188 participants ages 18-25 (US-based, Prolific), 373 conversations.

Metric	Result
Inter-human-expert agreement (Cohen’s κ)	0.45–0.64 (moderate)
LLM-vs-human-expert agreement (Cohen’s κ)	matches inter-human
Conversation-level evidence rate (skill-matched Executive LLM)	92.4% PM, 85% CR
Independent Agents (no steering)	significantly less skill-evidence
Telling subject “focus on this skill” alone	no significant effect (p > 0.6)
Creativity autorater vs human expert (n=180)	Pearson r = 0.88

This is the first scalable, ecologically-valid, psychometrically-controllable durable-skills assessment in the wiki — and arguably in the published literature given the validation rigour. Adds a row to ai-benchmarks.

Skills measured (Globerson et al. operational definitions)

Skill	Sub-skills	Conceptual basis
Collaboration	Conflict Resolution; Project Management	Team cognition (Cannon-Bowers & Salas); shared mental models; grounding (Clark & Brennan 1991)
Creativity	Generating Ideas (fluidity, originality, quality); Evaluating Ideas (elaborating, selecting); Building on Ideas	Said-Metwaly et al. systematic review; OECD 2019 fostering creativity report; CAT-aligned divergent + convergent thinking
Critical Thinking	Interpret and Analyze; Evaluate and Judge	Facione (Delphi report 1990); Ennis 1991; Halpern 1998

Notable methodological choice in the critical-thinking rubric: AI-Supported Exploration and AI-Supported Verification are explicit sub-categories. Globerson et al. treat AI fluency as a sub-component of critical thinking in 2026, not a confounder. A participant who blindly accepts the AI’s output is rated lower than one who uses the AI to fact-check, retrieve specific information, and stress-test arguments. This is a notable update on Facione/Ennis-era critical-thinking definitions.

Why this matters for the wiki

The durable-skills concept sits at the intersection of three open questions in the wiki’s labor-and-AI cluster:

1. What replaces the entry-level pipeline? Brynjolfsson et al. (2025) showed early-career workers (22-25) in AI-exposed occupations have ~13% relative employment decline since 2022, concentrated in automation uses. If durable skills are what differentiate non-substitutable workers, then operationalising them tells us which entry-level training/curriculum still pays — a question the wiki has not been able to answer concretely until now.
2. Which skills are decaying via ai-deskilling? Deskilling is the substitution direction; durable skills are the retention direction. The two concepts together carve up the labor-skill space: codified-and-AI-substitutable vs. open-ended-and-AI-resistant.
3. What’s the operational definition of “AI-augmentative” work? automation-vs-augmentation tracks the empirical split between AI uses that substitute for vs complement labor. Durable skills are the content of what augmentation augments — if we can measure them at scale, we can audit whether deployments are actually building durable-skill capacity in the workforce or hollowing it out.

Cross-source positioning (descriptive only)

Other wiki sources have touched the underlying construct without using “durable skills” terminology:

• Brynjolfsson, Chandar & Chen (2025) — Canaries in the Coal Mine. Empirical labor-market evidence that young workers in AI-exposed occupations are losing employment, but doesn’t operationalise which skills remain valuable. Durable-skills work is the missing supply-side measurement.

• AI Index 2025 / 2026 — repeatedly cite McKinsey workforce-reshaping statistics (“46% expect >20% of the workforce to need reskilling”) without a clean operational definition of what the reskilling targets should be. Durable skills are a candidate answer.

• Anthropic Economic Index (Q4 2025) — measures task complexity and AI autonomy at population scale on Claude.ai conversations. Doesn’t directly measure individual durable skills, but the “task success” primitive captures something adjacent.

• Anthropic Economic Index 5 — Learning curves (Feb 2026) — quantifies a different kind of learnable skill: AI literacy itself. High-tenure Claude users achieve ~3-4 percentage points higher task success than lower-tenure users after controlling for task type, model selection, and conversation complexity. Model selection slopes (Opus differentially used for higher-value tasks: +1.48 pp per +$10/hour on Claude.ai; +2.79 pp on the 1P API) demonstrate that choosing the right model tier is itself a learnable behaviour visible in usage logs. This is a complementary kind of “durable skill” — not the open-ended, contextually-situated skills the Globerson paper measures, but rather a procedural meta-skill (knowing how to deploy AI well) that nonetheless is learned-by-doing and produces measurable productivity returns. Open question for synthesis: is AI-literacy a 5th durable skill alongside collaboration, creativity, critical thinking — or is it a fluency that enables the other three to operate in AI-saturated work environments?

• Ransbotham et al. 2024 — Learning to Manage Uncertainty, With AI (MIT SMR × BCG) — provides the organizational-level counterpart to Globerson’s individual-level measurement. The 5-question organizational-learning battery (test-and-learn; failure tolerance; postmortem culture; learning codification; knowledge sharing) and the 4-question AI-specific-learning battery (AI generates new learning; AI learns from performance; human-feedback-loop AI; employees learn from AI) together define what kind of organizational capacity AI augments. Augmented Learners — the 15% combining both — outperform Limited Learners by 1.6× on uncertainty management and 1.4× on revenue benefits. The construct is adjacent to durable skills but at a different layer: durable skills measure individual humans; Augmented Learners measure organizational capacity. Both depend on the same underlying capability — converting tacit knowledge into surfaced, shareable, AI-readable form.

• Kiron & Schrage 2026 — How to Reap Compound Benefits From Generative AI — supplies the strongest claim in the wiki to date about which durable skill is now most economically valuable: the expert as evaluator. Direct quote: “The expert as evaluator is not a transitional role.” Mechanism: where verification compares output against existing standards, evaluation may generate standards that did not exist before — and that requires domain expertise. The column reframes durable expertise around a specific operational role (recognizing what “not perfect” actually means; knowing what iteration may reveal). Also names the Polanyi tacit-dimension breach — LLMs no longer just codify explicit knowledge; they infer tacit knowledge from behavioral traces at scale. Tacit expertise was a moat; it is no longer. This sharpens the durable-skill question: if AI inferences tacit knowledge, what’s left for humans? The answer the column proposes: the judgment to evaluate what AI produces, which itself depends on having previously developed deep expertise. Implication: deploy AI first in domains where your people already have deep expertise, not because AI needs hand-holding but because evaluation requires recognizing meaningful gaps. CLAUDE.md as a worked example: when developers write CLAUDE.md files specifying what “good” code looks like, that is tacit-to-explicit conversion at the team level — operationally identical to what Globerson’s Vantage assessment elicits at the individual level.

• OECD Digital Education Outlook 2026 (cited in Globerson et al. ref [68]) — mainstream educational policy framing of “21st-century skills” matches the durable-skills construct closely.

• Hansen) — supplies the construct from the educational-supply-side vantage. The wiki’s first institutional-leader source on open-education-as-AI-era-skills-investment. MIT OpenLearning’s Universal Learning architecture (22 modules of Universal AI live as of March 2026; 18 more verticals “cooking”; Universal Climate + Energy / Biology / Health following) is explicitly positioned as the supply-side investment in AI-era skills at scale. The 1-billion-learners-in-10-years ambition — “Aspiring to educate 1 in 8 in the world” — represents a 2x scale-up over OCW’s 25-year trajectory and is only achievable via AI-amplified distribution (translation into 12 languages launching May 2026; personalised education launching summer 2026; AskTIM AI tutor across all modules). Two-axis vocabulary contribution: personalized vs personalizable materials (the agent reorganises content for each learner versus the content has the wide-open structure that enables reorganisation — Newton’s distinction); and horizontal-knowledge architecture (AI+x verticals rather than discipline-by-discipline) anticipating that “the world is organized horizontally” not vertically. Convergent with Sternfels’ aspiration-setting / judgment / discontinuous-leap-thinking / human-to-human skill framework from the educational-supply side: Universal Learning’s AI + x verticals (AI + drug design, AI + satellite data, AI + law, AI + health) are precisely the AI-meets-domain skills the McKinsey hiring overhaul prioritises in candidates.

• Thompson 2026 (NYT The Daily) — supplies the construct in operator vocabulary from the developer-side. Thompson reports a clean inversion of what developers thought their durable skills were: “What developers are experiencing right now is something that maybe seems a little paradoxical, which is that they had spent years developing these very, very hard technical skills, and it turns out those are some of the easiest things to automate. The hard stuff to automate is like talking to our colleagues and our customers and figuring out what should we be building. Setting priority, setting strategy. AI can’t do that. There are still truly human skills.” The named-by-developers durable skills cluster reduces to: communication / spec-elicitation / priority-setting / strategy / deciding-what-to-build. Direct convergence with the Globerson collaboration / critical-thinking / creativity framework and with Sternfels’ aspiration-setting / judgment / discontinuous-leap-thinking / human-to-human framework — three sources from three vantages (Google Research psychometrics, McKinsey GMP hiring criteria, NYT journalism field-report) now name overlapping but non-identical sub-skill clusters. Synthesis target: map the three skill-vocabularies onto a common taxonomy.

• McKinsey) — supplies the construct from the hiring-criteria angle. Sternfels names four durable leadership skills models lack: (1) aspiration-setting (“They’re not good at setting the right aspiration level. Great leaders help an organization set the right ask: what should we aspire to?”), (2) judgment (“There’s not truth in the model. There’s not judgment in the model. Humans need to impose those parameters.”), (3) discontinuous-leap thinking (“They’re great at a linear approach to problem solving. What they’re not great at is discontinuous leaps, truly novel thinking.” — McKinsey is now consciously rebalancing toward liberal-arts undergraduate degrees), and (4) human-to-human skill (built via team sports / retail jobs / real-world group work). McKinsey applied 20 years of internal analytics to its own hires and surfaced three biggest predictors of partner-track success that mainstream credentials under-weighted: resilience-after-setback beats perfect marks; team-sport / retail-job experience under-weighted; aptitude-to-learn-novel-stuff beats subject-mastery — the firm now deliberately tests candidates in no-pattern-recognition assessment environments. This is the wiki’s first industrial selection-instrument validation of the durable-skills frame: McKinsey isn’t measuring durable skills in workers post-hoc, it’s measuring them in candidates pre-hire and changing what gets weighted. Direct convergence with the Globerson collaboration / creativity / critical-thinking framework, articulated in vendor self-narrative rather than psychometric form.

• Developer Intelligence team) — supplies the construct from the engineering-role-evolution vantage, with a structural model: the evolved T-shaped engineer with four skill domains. Vertical stem = deep core SWE specialisation (more important, not less); horizontal bar adds a new layer of AI engagement (steering, eval design, output assessment); and the wings extend into Adjacent Engineering (cybersec, privacy, deployment infra) and Adjacent Non-Engineering (business / user context). The Adjacent-Non-Engineering wing is the talk’s name for the same skill cluster Globerson et al. measure as collaboration / creativity / critical thinking, and the same cluster Thompson reports developers naming as communication / spec-elicitation / priority-setting / strategy, here packaged as the value-translator role: “engineers are increasingly acting as value translators, taking the user needs and business needs to precise requirements.” The talk’s one-line synthesis of what humans keep when agents take execution — “delegate tasks, not judgment” (used twice) — is the most quotable single articulation in the wiki to date of the durable-skill boundary. Two further contributions: (a) identity threat as a labelled phenomenon (citing DORA research) — “many developers are currently feeling a very real identity threat… we need to shift how we measure our worth” — the engineering-side correlate of the broader ai-employment-effects role-content shift; (b) explicit convergence claim across disciplines: “software engineering isn’t the only discipline getting more T-shaped — we are seeing this exact same evolution in our product managers, our UX researchers, and more” — the multi-role articulation of the durable-skills cluster’s becoming-the-shared-job-description trajectory. Same parent organisation as Globerson et al. (both Google Research-lineage); the two together form a paired engineering / psychometrics view of the same skill cluster.

• Storoni 2026 (HBR IdeaCast) — supplies a regulation meta-skill on top of the conventional durable-skills cluster: self-regulation under AI-era uncertainty pressure — specifically, the ability to stay at the top edge of gear-2 (slightly apprehensive, jittery, but self-controlled) without tipping into gear-3 panic. The neuroscience claim: when you encounter a situation more uncertain than you want it to be, your brain releases norepinephrine, which both causes the subjective tension and primes neural plasticity for learning. Workers who can stay at that top-of-gear-2 edge get the maximum learning benefit; workers who flee into avoidance or tip into gear-3 panic lose the priming. “Learning how to control yourself, regulate your brain from tipping into gear three … that is the state that, if we can train ourselves to embrace it, will last us and take us far through this era of rapid change.” This is not a fourth durable skill alongside collaboration / creativity / critical thinking — it is the operating condition under which the conventional three become exercisable. Storoni’s other contributions to the durable-skills picture: the “we have now effectively all become AI managers” framing (durable skill = working at the management/judgment layer, not the tool-execution layer), the intrinsic-motivation / learning-progress mechanism (durable engagement under conditions where external scaffolds have failed), and the counter-intuitive boredom fix (oversight work — increasingly common as AI takes execution — needs harder tasks and more feedback channels, not simpler ones). Per Lifecycle confidence rules this single-source popularisation contribution does not lift concept confidence; its value is expanding the meta-layer of the durable-skills picture rather than confirming the empirical cluster.

• Egon Zehnder) — centers adaptability as the primary durable skill of the AI era, with the rhetorical move “Adaptability is the new resilience.” Three layers of operationalisation: (i) how leaders lead — continuously restructure agile teams, harness diverse perspectives, nurture “creative abrasion” (Linda Hill), bring the outside in; (ii) how they hire — reframe job descriptions from tasks to challenges, look for candidates who have adapted before, interview for what new skills they have learned and what they have recently changed their mind about; (iii) how they build their own muscle — seek minor discomfort, reframe situations, practise adapting. Empirical anchor: Egon Zehnder survey of 1,200+ global CEOs; 92% agreement with the statement that adapting “beyond anything I’ve previously imagined” is now required. Source-quality caveat: this is HBR.org Partner Content (sponsored / advertorial) — the survey methodology is not disclosed, the framing strategically positions Egon Zehnder’s leadership-advisory offering, and per the wiki’s Lifecycle vendor-source rule this source does not raise the concept’s confidence. What it does add: a centerpiece treatment of adaptability that the wiki’s other sources mention only in passing, and a direct hiring-criteria convergence with Sternfels above (both 9 Feb 2026; Sternfels names “aptitude-to-learn-novel-stuff”, Ross-Schneider names “interview for what new skills they have learned” — same construct from two consulting-firm self-narratives on the same date).

The construct is recurring; the operational measurement is new in 2026; the vendor-narrative centering of adaptability specifically is new in early 2026 (Ross-Schneider).

Open questions and limitations

• Cultural situatedness. Globerson et al. explicitly acknowledge: “human skills are culturally situated, and will therefore also focus future work on exploring performance across diverse cultural settings and languages to ensure our technology is inclusive and equitable.” Current empirical results are US, English-speaking, ages 18-25. Generalisation is a real open question.
• Construct validity vs criterion validity. The Vantage validation establishes construct validity (the test measures the construct as defined by the rubric). It does not establish criterion validity (the score predicts real-world outcomes — manager reports, teacher reports over semesters, downstream career success). Criterion validity is the gold standard but is “largely impractical” to collect at scale, per the paper.
• Single-source operational definition. The wiki has only one source operationalising the durable-skills construct so far. The terminology cluster (durable skills / 21st-century skills / future-ready skills / human skills) is consensus across the broader literature, but a 2nd ingest would substantially raise confidence and surface vocabulary disagreements.
• Are durable skills actually durable under sustained AI augmentation? Brynjolfsson, Li & Raymond (2025) showed augmentation under deliberate human-in-the-loop design produces +15% productivity and equalising effects — but also small quality declines among top performers. If the AI’s training data degrades because skilled humans lean on it instead of producing original judgments, do “durable skills” stay durable? Open question.

Human skills growing twice as fast, and evaluating a fleet as the terminal skill ( WEF-PwC, July 2026 + Brynjolfsson, August 2026)

Peter Brown supplies this page’s most direct demand-side measurement, from the WEF–PwC entry-level report: in AI-exposed sectors, demand for human skills is growing at two times the rate — “they’re human skills: judgment, problem solving, critical thinking, relationship building.” His framing of the mechanism is the one that makes it a durable-skills claim rather than a hiring-fashion claim: in entry-level roles “much of the tasks they’ve done traditionally can be automated — but what can’t be automated is actually the skills that are in huge demand.” He names the second misconception explicitly: “that AI automatically means that human skills become less important — and actually we’re seeing the reverse. As AI is getting more advanced, the requirement for more advanced human skills is increasing.” His three-part composite for young people is domain expertise + AI fluency + deliberately accumulated human skills, with the caveat that the last “is really easy to say, hard to do.” (Second-hand from a report co-authored by his own firm; see that source’s quality note.)

Brynjolfsson declines to name skills at all when asked what children will need, and gives a framework instead — which is the more interesting answer for this page:

“My framework is less to go into specific skills and more into this broader framework… having that intention, figuring out what is it that you really want to do. I think that’s a skill that can be engendered, even taught… because I think almost everyone is going to be managing not just an agent, but a fleet of agents. They’ll be like the CEO of their own little entity, and they’ll have to have those leadership skills leading agents to direct them — and the ones who are good at pointing them in the right direction and then evaluating them are going to really thrive.”

Two things to hold. First, the terminal skill he names is evaluation, not direction — which converges with the verification-is-the-bottleneck finding running through the CS329A lecture and Frey’s verification tax, and makes evaluation the one competence three quite different sources independently place at the centre. Second, the fleet-management framing matches Miller’s third hiring criterion and her 34-agent org chart, reached from advisory practice rather than economics — the same conclusion from two directions within ten days. His reframe of the acronym is the compact version: “when you hear AI, you should not think artificial intelligence. You should think amplifying intention.”

Debates and supersession

Expert generalists + renaissance developer + the Anthropic-competition exhibit ( AWS London Exec Forum 2026)

Allen’s keynote names two vocabularies adjacent to durable skills at AWS-Executive-in-Residence altitude: expert generalists (Martin Fowler’s term, recommended by Allen) and renaissance developer (Werner Vogels’ final re:Invent 2025 keynote framing). Both describe the shape of the worker whose skill profile survives the agentic transition: “domain experts pushing outward; generalists deepening their domain expertise” — the meeting point in the middle is the renaissance developer / expert generalist.

The empirical anchor Allen offers is the Anthropic competition exhibit (~17:55–18:30): in a recent Anthropic-run agent-building competition, 1st place went to a lawyer, 2nd place to a cardiologist from Poland (who built an AI-agentic patient-care platform that guides patients after they leave the doctor’s office, “flying around the world with just Wi-Fi to build this”), 3rd place to another cardiologist. “No developer in the top three.” Allen’s reading: “Think in your business: who has the domain knowledge?”

The wiki’s reading: the Globerson et al. durable-skills measurement battery (collaboration, creativity, critical thinking, contextual reasoning) is necessary but not sufficient to describe the renaissance-developer skill profile. The Anthropic-competition exhibit adds the dimension of domain expertise × tool-fluency as the differentiator — durable skills enable judgment under uncertainty, but the competitive advantage in the agentic era goes to those who can pair deep domain expertise with goal-based-agent tool-fluency. This is the AWS-vendor-altitude framing of what durable skills compose with to win in the agentic era — a complement to the supply-side measurement question Globerson et al. answer.

Primary source for the term (added 2026-06-25). Allen cited expert generalist secondhand; the wiki now holds its anchor — Joshi, Venkatraman & Fowler (2025), promoted to its own expert-generalist concept page. Fowler’s six characteristics (curiosity, collaborativeness, customer-focus, favoring fundamental knowledge, a generalist+specialist blend, sympathy for related domains) are the software-practitioner articulation of the durable-skills cluster, and the article’s LLM thesis — generalists who hold the fundamentals extract more from LLMs because they interrogate rather than accept output — is the same experience-beats-AI-native finding from the developer’s chair. The article concedes the Expert Generalist skill is hard to assess at scale (“often requiring intensive participation from known-capable Expert Generalists”), which is the live tension with Globerson et al.’s scalable-measurement programme.

The MGI Skill Change Index as a labour-market-data-grounded systematic measure ( MGI 2025)

MGI’s Agents, Robots, and Us (November 2025) introduces the Skill Change Index (SCI) — the wiki’s first labour-market-data-grounded systematic measure of which skills will change most and least under AI automation. Three properties make the SCI a useful complement to the wiki’s prior durable-skills vocabularies (Globerson et al. assessment methodology, Storoni’s neuroscience-vantage self-regulation-under-uncertainty, Schoening’s agency-as-durable-skill):

(a) Coverage: 6,800 employer-cited skills from Lightcast across ~1,800 US occupations. Not a hand-selected list of “AI-era skills” — every skill with >5% job-posting prevalence in the US labour market. The breadth makes the SCI a systematic baseline rather than a vantage-specific framing.

(b) Time-weighted decomposition methodology. Each skill is scored on the fraction of associated time spent on automatable work activities, weighted by occupation employment. Skills with 55%+ time-spent in non-automatable work activities are classified people-led; 55%+ in automatable are AI-led (further divided into agent-led / robot-led depending on physical-capability requirement); all others are shared. The classification is time-spent-weighted, not capability-binary — a skill can be people-led even if some of its hours overlap with AI-capable tasks, as long as the dominant usage is in non-automatable work. The mapping uses OpenAI GPT-4o for ~3.4M occupation-DWA-skill labellings with a manual 1,000-cell template for validation + iterative prompt refinement.

(c) Scenario-bracketed forecasts. Two adoption scenarios for 2030: midpoint (the report’s primary lens) and early (faster diffusion). The SCI quartile values differ markedly across scenarios — midpoint top-quartile 33% exposure vs early-scenario top-quartile 59% — making the uncertainty band transparent. The wiki can therefore cite which skills are durable under both scenarios (the bottom-quartile and lower-middle-quartile skills) with confidence-graded weight.

The SCI’s three skill-evolution paths:

1. Top quartile (highly exposed → will decline in demand): specialised skills like specific programming languages (SQL programming), accounting processes, invoicing, basic technical knowledge.
2. Middle quartiles (will evolve in nature and application, not rise or fall): transferable skills + AI fluency itself is in this band — communication, customer relations, detail orientation, management, writing, quality assurance.
3. Bottom quartile (will endure): skills grounded in human connection and care — leadership, coaching, negotiation, mentorship, healthcare skills, good driving record.

By skill category (Exhibit 13, share of skills in each SCI exposure band):

Category	Low (durable)	Moderate	High (declining)
Assisting and caring	54%	36%	10%
Communication, collaboration, creativity	28%	58%	14%
Management skills	29%	49%	22%
Handling and moving	29%	42%	29%
Constructing	25%	44%	31%
Working with machinery + equipment	18%	55%	27%
Information skills	18%	53%	29%
Digital skills	11%	47%	42%

The assisting-and-caring category is the wiki’s most-protected durable-skills cluster (54% low-exposure, only 10% high-exposure); the digital-skills category is the most-exposed (42% high-exposure). The pattern matches Globerson et al.’s collaboration / creativity / critical-thinking measurement battery being the assessment-methodology for moderate-exposure skills, while ai-employment-effects data shows moderate-exposure skills will evolve, not decline.

Eight high-prevalence transferable skills that endure across the workforce, with explicit splits between what humans vs agents will do (Exhibit 11):

Skill	Relevance %	Agents/Robots will	People will
Communication	99	Generate content, accelerate data flow	Refine nuance and storytelling
Management	94	Automate scheduling, monitor metrics	Coach and lead hybrid teams
Operations	84	Execute routine tasks, optimise efficiency	Design smarter processes and strategise
Problem-solving	83	Identify patterns, propose options	Interpret findings, make judgments
Leadership	83	Drive change, support decision-making	Guide and motivate teams
Detail orientation	80	Run quality checks, flag anomalies	Audit outputs, validate outcomes
Customer relations	80	Route requests, handle routine queries	Strengthen loyalty, build relationships
Writing	76	Produce drafts, propose revisions	Refine text, craft story

The framing — AI does the quantity, humans do the quality — is convergent with Storoni’s neuroscience-vantage reframe of efficiency (gear-3 quantity / gear-2 quality) now at structural-modeling altitude.

The SCI is the wiki’s strongest structural-prospective anchor on which skills are durable. Joins Globerson et al.’s measurement-methodology layer (how to test for durable skills), Storoni’s neuroscience layer (what makes a skill durable under uncertainty), Schoening’s personal-agency framing (which mindset-traits compound into durable skills), and Forsgren & Macvean’s engineering-team-core-skills vocabulary (which 10 capabilities matter for AI-era developers). The five-source convergence on the durable-skills concept is now multi-altitude — measurement + neuroscience + agency + engineering-team + structural-modeling.

The Coursera 120% YoY critical-thinking enrollment growth demand-side anchor + Kocherlakota end-to-end systems thinking prescription ( WP Intelligence May 2026)

Giles (May 2026) supplies the wiki’s first demand-side empirical anchor on the durable-skills market — measured at learner level, not employer-side:

“In its 2026 Job Skills Report, online learning platform Coursera surveyed more than 6 million learners across nearly 7,000 organizations and reported a 120 percent year-over-year increase in enrollments in courses on critical thinking.”

A 120% YoY enrollment jump in critical-thinking courses across a 6M-learner sample is the wiki’s strongest single-line demand-side signal that durable-skills training is materialising as an active labor-market behaviour rather than only a prescriptive recommendation. Convergent with the MGI Skill Change Index AI-fluency-demand-grew-6.8×-in-two-years finding on the supply-side / employer-posting end — the Coursera data is the workers’-side complement (learners actively enrolling rather than employers actively posting).

The accompanying executive-altitude end-to-end-systems-thinking prescription Giles surfaces, from Swamy Kocherlakota (Zscaler, agentic-AI lead): “A lot of things companies currently do in silos will be collapsed because of AI. That’s why the skill set that’s going to be most needed is end-to-end systems thinking.” + (separately) “If you are working like a robot, your job will be taken by a robot.” The wiki’s clearest single-source articulation of the what-skill-survives-the-silos-collapse question. Convergent with the existing Globerson-et-al. measurement framework (the four sub-skills include AI-supported critical thinking and novel reasoning — sister concepts to systems thinking) and with Storoni’s gear-2 neuroscience framing (the cognitive state in which systems-thinking is even possible).

Plus the UiPath case-study-vs-online-training worked example Giles cites (Agi Garaba, UiPath CPO): “online training … is useful for helping people understand the fundamental concepts of softer skills. But she has found the most effective way to develop those skills is to bring groups of employees together to work on actual case studies — and to listen to UiPath leaders sharing relevant personal experiences.” The wiki’s first practitioner-altitude statement that softer-skills training is more effective via case-study practice + leader storytelling than via abstract online courses — a pedagogical finding worth tracking against the durable-skills measurement literature (Globerson et al.). The pairing implies: measurement methodology + case-study-based pedagogy may need to co-evolve.

Upskilling over hiring: the institutional-knowledge case (LF Global + LF Europe 2026)

The two 2026 Linux Foundation State of Tech Talent reports (Marco Gerosa & Adrienn Lawson) reframe durable skills from a measurement problem into an organisational-strategy one: given an AI-era skills gap, do you upskill existing staff or hire externally? The survey answer is emphatic and quantified:

• Upskilling is the primary response (57% global / 63% Europe), ahead of external hiring, rated important by 94% of organisations — and explicitly favoured over hiring: 7.9× for business context, 7.7×/5.6× for retention, 7.3×/6.3× for team cohesion, 5×/5.8× for lower total cost (global / Europe). Organisations are 3.5–3.7× more likely to upskill than hire across strategic domains.
• Why: institutional knowledge is a durable asset. The reports’ framing — “upskilled talent doesn’t take wing and fly, they grow roots” — treats upskilling as institutional-knowledge preservation, not just gap-closing. External hiring is slower and riskier: new hires take 53% longer to reach productivity and 23% resign within six months. This is the workforce-economics complement to the durable-skills thesis: the skills worth measuring are also the skills worth retaining and deepening in place.
• Skills-over-credentials at the hiring gate. Degrees/titles/years-of-experience are “proving insufficient” as roles change faster than those signals; certifications (important to 76% of hiring managers) and demonstrated job-relevant skills are the rising trust signals — and technical training (93%) ranks above compensation (91%/83%) as a retention strategy. (Interest-alignment caveat: the publisher sells certifications; see the source pages.)
• The pipeline-investment corollary (Europe report): with entry-level technical roles contracting (−3% in Europe), upskilling becomes not only gap-closing for current staff but a pipeline investment for the mid-to-senior talent organisations will need — tying durable-skills development directly to the ai-employment-effects entry-level-decline finding.

This sits alongside the page’s measurement spine (Globerson et al.) and the MGI Skill Change Index: measure which durable skills matter, then develop them in-house as the dominant organisational strategy.

The mindset-not-skillset inversion ( HBR June 2026)

Marco Argenti (CIO, Goldman Sachs) supplies the wiki’s sharpest counter-framing to the skills-inventory approach this page otherwise documents. The standard durable-skills move is to identify and retain the skills AI can’t substitute. Argenti argues the framing of preservation is itself the trap: when a banker asks “what’s the 10% of my job AI will never do?”, his answer is “let go of that 10%.”

• The horse-rider→car-driver analogy. None of the specific skills transfer from riding to driving; what transfers is the reflexes and instincts — and the durable human residue Argenti names is instincts, judgement, and values, not any task-skill. This aligns with the wiki’s measured durable skills (critical thinking, collaboration, judgement) but reframes them as a substrate that survives role-metamorphosis rather than a list to defend.
• From skillset to mindset. “Don’t just reskill, reimagine skills and build new habits.” The durable capability is the capacity to let trusted habits die and adopt a new professional identity — a meta-skill above any catalogued competency. The hardest part, Argenti notes, is personal and cultural, not technical.
• Convergence with the wiki’s judgement spine. This restates, from a first-party CIO vantage, the McKinsey claim that judgement and human-to-human skill are what models lack — and operationalises it as a posture (operator → supervisor/mentor) rather than a training curriculum.

Rising skill thresholds + AI fluency as a tenure-complement ( Emerson, Kropp et al. 2026)

BCG’s labor-disruption report turns the durable-skills thesis into a workforce-economics claim: as AI absorbs routine execution, skill thresholds rise — “redesigned roles will demand greater expertise, oversight, and accountability… the more durable roles tend to require higher credentials and greater seniority.” Durable skills aren’t just what survives; they become a barrier to entry (higher credential bars, transition friction) even where headcount holds. Two BCG specifics sharpen the page:

• AI fluency becomes a complement to tenure. Readiness for augmented, higher-order roles is assessed increasingly on whether you proactively adopt AI tools and rethink the work — sometimes advantaging AI-fluent juniors and recent graduates over more-experienced staff. This is the demand-side mirror of the Coursera 120% critical-thinking-enrollment signal and complicates the simple “seniors win” reading: the durable bundle is judgment + AI fluency, and fluency is age-agnostic.
• Frequent, not one-time, upskilling. BCG’s CEO directive is that workers “will likely require more frequent (versus one-time) upskilling as the technology evolves,” with structured redeployment pathways per role-segment — making durable-skills development a continuous organizational function, not a one-shot reskilling event (echoing the Linux Foundation upskill-over-hire finding).

Three June-2026 practitioner statements: expertise + taste as the lever on commoditised model intelligence (Ng, Mollick, Chou)

Three mid-June-2026 video interviews land the durable-skills thesis from three practitioner chairs and converge tightly with the page’s judgement spine (Sternfels, Argenti, Thompson) and the empirical AEI returns-to-expertise report:

• Ethan Mollick — experience beats AI-native youth; taste is the differentiator. Mollick’s headline directly inverts the digital-native intuition: young people are not “AI natives” (they’re “conduits to Claude” who can’t judge the output), and he cites the same BCG study where junior employees were worse at using AI. As model output commoditises (“the death of the movie star”), the residue is taste — variation, point of view, the director’s hand — possibly “the most valuable skill of the AI era.” He also relocates the binding constraint to evaluation speed (an expert skill), echoing the page’s expert-as-evaluator claim (Kiron & Schrage), and supplies the jagged-frontier vocabulary the page already uses (jagged-frontier).
• YC — expertise is what you bring to boundless intelligence. “You need a certain amount of expertise to know what to do with this boundless intelligence imbued in the model”; a decade-plus of domain experience is what turns a general model into something “world class.” The age-of-the-40-year-old-solo-founder framing makes experience + taste the lever (“you don’t have to be 40, you just have to have taste”), the founder-chair instance of Argenti’s instincts/judgement/values survive role-metamorphosis.
• Andrew Ng — the scarce input is deciding what to build. Ng’s product-management bottleneck becomes an everything-bottleneck and his small teams of high-context, empowered generalists relocate value to scoping, judgement, and cross-functional context — the same skill cluster Thompson’s developers name (communication / spec-elicitation / priority-setting / deciding-what-to-build).

The three add no new measurement methodology (so they do not lift the concept’s confidence past its 0.95 ceiling), but they are the strongest practitioner-vantage convergence in the corpus on the page’s central claim: as execution commoditises, the durable bundle is domain expertise + judgement + taste, and (per BCG) the fluency to wield AI well rides alongside it.

A fourth June-2026 voice, Sridhar Ramaswamy (CEO, Snowflake), names the meta-skill side directly as a hiring criterion: malleability/adaptability — “tell me something meaningful you have changed in yourself over the past year” — plus don’t-set-limits / first-principles generalism (“I’m an engineer by training but I think from first principles about everything from communication to marketing to strategy to sales”) and curiosity. This is the Argenti mindset-not-skillset / capacity-to-change-professional-identity claim restated as an interview question, and convergent with the Ross & Schneider adaptability-is-the-new-resilience centering and Sternfels’ aptitude-to-learn-novel-stuff hiring shift.

Cognitive coverage + token-efficiency as premium skills in token abundance ( Possible, June 2026)

Nadella (June 2026) names two durable capabilities specific to the token-abundance regime, both convergent with the page’s judgement/evaluator spine:

• Cognitive coverage. “Just like test coverage, whenever an agent does work it creates a quiz so we humans learn what it did.” The durable skill is the deductive understanding and evaluation of agent output — “the expertise is always in abundance; it’s your cognitive coverage of that expertise that’s at a premium.” This is the platform-CEO operationalisation of the expert-as-evaluator claim (Kiron & Schrage) and Argenti’s everyone becomes a manager, framed as a measurable practice (a quiz per agent task) rather than a posture.
• Token-efficiency + eval/rubric authoring. “Whoever figures out they can use tokens more efficiently for an outcome that matters is going to get ahead by definition.” The premium skill is setting “the rubric and eval dimensions such that they capture the high taste only you can define” — and knowing the system’s limits well enough to apply “don’t use frontier models for non-frontier problems.” This adds a meta-skill to the durable-skills cluster: the judgment to spec objectives/evals and allocate model-tier-to-task — adjacent to the AEI AI-literacy-as-learnable-meta-skill finding, here named from the platform-CEO vantage as a source of durable advantage in an era where raw intelligence is cheap.

The cognitive-science floor: why the residual is durable ( Stanford GSB, June 2026)

Most durable-skills sources are labour-market or practitioner evidence about which skills survive. Douglas Guilbeault (Stanford GSB) supplies the wiki’s first cognitive-science argument for why a residual exists at all. His distinction — humans satisfice (Herbert Simon: do a lot with a little, under constraint) where LLMs optimize (brute-force statistical prediction over massive data) — grounds the durable cluster in specific capacities AI’s optimization approach may be structurally unable to reproduce: conceptual leaps / insight (a jump from a random to an ordered state, not a smooth within-distribution step), metaphor and analogy (see analogical-reasoning), an aesthetic “vibes” / taste sense, and making meaning from genuine disorder. This is the theoretical underside of Mollick’s taste/experience/judgment edge and Argenti’s instincts-not-skills — and a caution against the “humans are just prediction machines” framing that, Guilbeault warns, leaves people unable to “defend their relevance.” It also connects the durable-skills supply-side question to the jagged-frontier capability-ceiling question: the skills that are durable are precisely the ones on the rough side of the frontier for a reason.

The 5 C’s — a book-length popular vocabulary for the cluster (Wood, WorkLab, June 2026)

Aneesh Raman (LinkedIn) names the durable-skills cluster creativity, curiosity, courage, compassion, communication — the “5 C’s,” offered as the intersection of “EQ and IQ” and “consciousness and conscience.” Two framing choices distinguish this from the page’s other vocabularies:

• Explicitly trainable, not innate. Every C is framed as a “muscle” buildable via deliberate practice, citing neuroplasticity and a (secondhand, unverified) account of Mozart’s perfect pitch as learned rather than gifted — the same skills-are-trainable move Ross & Schneider make for adaptability specifically, generalised across all five.
• An explicit priority ranking. Curiosity is named the single most urgent skill right now (“be curious about these tools… let that curiosity build”); creativity is named the best long-term bet, on the argument that as execution commoditises, humans increasingly need to generate new ideas/business lines rather than just execute — convergent with Mollick’s taste framing and Guilbeault’s cognitive-science case for conceptual leaps/insight as AI-resistant.

Courage and compassion are new labels in the wiki’s durable-skills vocabulary cluster — courage overlaps with Ross & Schneider’s adaptability (both name the willingness to act under uncertainty) and compassion is closest to the page’s collaboration construct (Globerson et al.) but foregrounds the interpersonal-care dimension more explicitly than “collaboration” does. Per Lifecycle rules this single-source, promotional-book popularisation does not lift the page’s confidence past its 0.95 ceiling — its value is a fifth independent vocabulary (after Globerson’s measurement battery, Storoni’s neuroscience, Schoening’s agency framing, and Forsgren-Macvean’s engineering-team taxonomy) converging on the same underlying cluster.

The ETS five-skill framework + an accredited degree built around it ( Silicon Valley Girl, July 2026)

Khan Academy CEO Sal Khan names a sixth independent durable-skills vocabulary — ETS’s framework: communication, collaboration, creativity, critical thinking, plus sub-frameworks such as leadership — and, distinctively among the wiki’s durable-skills sources, is building an accredited degree program around it rather than just naming it. The Content Institute (a joint venture with TED, where Khan succeeds Chris Anderson as “vision steward,” and ETS) is an accreditation-pending, competency-based bachelor’s/master’s degree capped at $10,000, assessed via group simulations and portfolios rather than GPA or transcript signaling — students work in Zoom-based teams to build a business (prototype, ad, business plan, competitive research) over several hours, then receive peer and multi-rater feedback on communication, collaboration, and creativity specifically. Six corporate co-design partners at launch (McKinsey, Bain, Google, Microsoft, Replit, Accenture) help define what “job-ready” durable-skills evidence should look like and commit to weighting Content Institute portfolios comparably to elite-university signals in hiring.

Two framing points connect this to the page’s existing vocabulary cluster:

• Vocabulary convergence. ETS’s five skills overlap heavily with Wood’s 5 C’s (creativity, curiosity, courage, compassion, communication) and with Globerson et al.’s measurement battery — a sixth independent naming of the same small human-skills cluster, this time from a psychometric-testing organization (ETS, the TOEFL/GRE publisher) rather than a labour-economics or neuroscience source.
• Assessment-as-product, not just assessment-as-measurement. Where Globerson et al. and Vantage are measurement instruments, the Content Institute packages durable-skills assessment directly into a credentialing product — closer in kind to a college degree than to a benchmark. This is the wiki’s first durable-skills source where the measurement methodology is the deliverable rather than a research contribution.

Khan’s explicit motivation ties back to ai-employment-effects: “Someone today graduates from Stanford… probably smart, good GPA… but that’s only a signal, not an evaluation.” The Content Institute’s pitch to employers is that 20-50 rigorously-designed, peer-reviewed project simulations are a stronger evaluative signal than a GPA-and-résumé screen — explicitly positioned as a lower-cost, non-elite-university alternative signal, not a replacement for the underlying skills claim already established on this page.

AI fluency as a universal, non-level-specific career-ladder expectation ( Netflix CPTO, July 2026)

Elizabeth Stone (Netflix CPTO) supplies the wiki’s clearest single-firm account of how a large incumbent operationalises AI fluency as a durable-skills expectation rather than a tool checklist. Rather than specifying what AI changes at each career-ladder level, Netflix layered an aspiration overlay across all talent — an experimentation mindset, judgment about where AI is and isn’t useful, and comfort with change — applied uniformly from new hires to the most senior leadership, “even if we’re not writing code as part of our day jobs.” This is a practitioner-HR-mechanism instance of the same claim BCG’s AI fluency becomes a complement to tenure finding makes structurally: the skill is level-and-age-agnostic, layered on top of (not replacing) role-specific competence. Netflix operationalises the expectation concretely in hiring: interviews probe how candidates think about AI/technology, and coding interviews now explicitly permit AI-tool use, “because that’s going to be part of what the work requires now.”

Stone also names systems thinking — see systems-thinking for the fuller treatment — as the specific content of the rising skill cluster at Netflix, giving this page’s cross-vocabulary convergence (Globerson’s collaboration/creativity/critical-thinking, Raman/Wood’s 5 C’s, ETS’s five skills) a seventh independent naming, this time from a for-profit streaming incumbent’s CPTO rather than an educator, consultant, or researcher.

The cognitive L1 cache, and three interview criteria for the agent era (Collison, July 2026 + Miller, August 2026)

Two sources a fortnight apart supply a latency argument for retaining knowledge in-head and a hiring argument for what replaces it as the scarce input.

Patrick Collison gives this page its most mechanical defence of knowing things. Borrowing Jeff Dean’s numbers every programmer should know, he applies the memory-hierarchy framing to human cognition: asking an agent to retrieve something “is a hell of a lot slower than knowing it in cognitive L1 cache. And you can have way more round trips in your brain than you can, you know, muttering through super whisper or typing it out.” Even granting the full capabilities of the models, “for a long time to come, neuronal lookups will be much faster.” His second argument is revealed preference rather than theory — whether at Stripe or at the frontier labs, “there still seems to be an enormous premium on cognitive ability,” so “renouncing that before there’s evidence that we’ve saturated those benefits would be premature.” The framing is useful because it does not depend on any claim about what models cannot do: it is an argument from round-trip latency and the number of iterations a mind can run against itself, which holds even if model quality is stipulated to be perfect. He reports keeping writing in-cache for himself — “I’ve yet to send… pre-written suggestions… I think I’ve still sent zero of those in my life.”

Allie K. Miller names the three criteria she now interviews for, and is explicit that the interview process itself is the thing most organisations have failed to update (“if you have not updated the way that you interview people in the last year, you are very behind”):

1. High agency — “are they going to let the AI dictate their day? And if you didn’t get what you want, are you just going to walk away and complain and still do things manually, or are you going to go in bulldozer mode and try it from 17 different angles?”
2. A strong sense of wonder — with the explicit disclaimer that it has “nothing to do with age, nothing to do with department… some people are two years from retirement who are the most curious people I’ve ever met.”
3. Systems thinking — “that’s been the really big change the last year, where we’re managing multi-agent systems.”

The third is where her argument bites, and it inverts the usual seniority story on this page. Because the skill in demand is running “a mini org, a temporary digital workforce of AI agents,” the people advantaged are those who have managed people before — and early-career ICs are “in a really tough position on that systems thinking side, very strong on the AI side.” This converges with Stone’s systems-thinking hiring criterion three weeks earlier, reached from an operator vantage and for a different reason (designing across business domains rather than supervising agent fleets) — two independent routes to the same criterion. Neither source lifts the page past its 0.95 cap: Collison’s is an unaudited revealed-preference claim and Miller’s is an advisory practitioner’s hiring heuristic.

The primary statement of creative abrasion, and collaboration as the ten-year skill ([[2026-08-16-hill-bloomberg-leaders-ceo-skills-age-of-ai|Hill / Bloomberg Leaders, August 2026]])

This page has carried Linda Hill’s vocabulary second-hand since the Ross & Schneider entry above, which cites her four adaptable-leader practices and her creative abrasion coinage without her being an author. Her August 2026 Bloomberg interview is the primary statement, and it promotes her to an entity page.

Creative abrasion, in her own words, is an organisational capacity a CEO should audit: “do you have enough of what’s referred to as creative abrasion? Do you have enough diversity of thought? And do you know how to embrace that diversity of thought, deal with that conflict? You need to amplify difference. But many CEOs minimize difference.” The mechanism: “everybody has a slice of genius… if you have very talented, passionate people and you let them say what they think, they’re not letting go easily. So you as a leader have to take those talented, passionate people — if you’ve unleashed their ideas, now you have to leverage and harness them to make sure you get some good solution out of it.” Her source for the hiring corollary, an early Google infrastructure leader: “you’re hiring them for their ideas, and if you don’t get those ideas, then shame on you… you can never have too many cooks in the kitchen.”

Two additions this page did not previously hold:

• Contextual intelligence, offered as a partner to emotional intelligence and framed as an attribution-error correction: “usually our context is shaping our individual behavior. Too often we think, oh, that person’s a good person, bad person, resistant, whatever. No. What is it about their context that is leading them to prioritize in ways that aren’t suited to what the company needs? Or are they constrained in some way that if I relieve that constraint, they begin to behave in ways the company really needs?”
• Empowerment as an overrated practice, which sharpens a word this cluster uses loosely: “empowerment simply means that you’re sharing with people the potential to have influence. So whether or not they get to exercise that potential is your call… I think sometimes people [say] ‘I empowered them.’ Well, in the end it’s on you.”

Asked directly which leadership skill will matter more in ten years than today, her answer lands squarely on this page’s central cluster: “your capacity to collaborate.” And on formal authority as a substitute: “I think formal authority as a source of power is actually quite limited in many ways in what you need to do now to influence people to get the job done.”

Confidence unchanged. Nothing in the interview is presented with a method, a sample or a result; its value is that constructs the page previously held at one remove now have a primary source. See also strategic-foresight for her wayfinding vs. pathfinding distinction, which is the more novel contribution of the same episode.

Related concepts

Related concepts

• expert-generalist — the software-practitioner articulation: Fowler’s trait-set is the developer-chair version of the durable-skills cluster.
• jagged-frontier — the capability-ceiling counterpart: durable skills cluster on the rough side of the frontier; Guilbeault argues the cognitive reason the two line up.
• ai-deskilling — the inverse: which skills are getting hollowed out by AI use.
• ai-employment-effects — the labor-market consequences. Durable skills are the supply-side answer to “what protects employment?”
• ai-benchmarks — durable-skills assessment is methodologically a new type of benchmark (for humans, not models).
• automation-vs-augmentation — durable skills are the content of augmentative work; codified knowledge is the substrate of automative work.
• generative-ai — the substrate that creates the deskilling pressure and powers the new measurement methodology.
• responsible-ai — the cultural-situatedness limit and the AI-supported-critical-thinking sub-skills are RAI-adjacent considerations.

Management as the durable skill in agentic work (added 2026-08-30)

Ryan Carson makes the strongest version of this page’s claim from inside agentic software work: “the reason why I think folks of a certain level of experience are ready to cook right now with AI is all those management skills that you have developed over your career… How do you structure? How do you team? How do you prioritize? How do you delegate? Where do you micromanage? Where do you let go?” Notable because it is an experience-as-advantage claim in a domain where the randomised evidence (Cui et al.) shows juniors gaining most from AI assistance — the two are compatible only if the durable skill is coordination rather than production. See agent-fleet-management.