The half-life of a technical skill just collapsed from a decade to a fiscal year, and every hiring process still screening for accumulated skills is buying inventory that expires before onboarding ends. To be precise about what decayed: fundamentals — systems thinking, statistics, writing, judgment — hold their value fine. It's the operational layer that now rots in months: the specific tools, model APIs, frameworks and workflows that job descriptions obsess over. The prompt patterns of eighteen months ago are archaeology; the agent stack of last quarter is already being rebuilt. When the thing you screened for expires this fast, the only durable signal is the rate at which someone acquires the next thing. Stop hiring for the skill. Hire for the slope.
What decays fast, what compounds
| Layer | Examples | Half-life | Hiring implication |
|---|---|---|---|
| Tool-specific | A vendor's API surface, prompt idioms, this quarter's agent framework | Months | Never a screening criterion — teach it in week one |
| Workflow | How to structure evals, agent orchestration patterns, AI-assisted dev practice | 1-2 years | Screen for having built one, not for a specific stack |
| Craft fundamentals | Systems design, statistics, writing, debugging, security instincts | 5-10 years | Screen hard — this is what makes new tools learnable |
| Meta-skills | Learning speed, judgment under ambiguity, taste for what matters | Career-length | The primary hiring signal; only visible in proof-of-loop |
Hire learners with proof-of-loop
- Proof-of-loop is concrete evidence of a recent learning cycle closed end-to-end: encountered unfamiliar territory → acquired the capability → shipped something real with it → can articulate what they'd do differently.
- Ask for the last 90 days, not the last decade: 'What did you learn recently that you didn't need before, and what did it ship?' Strong candidates answer instantly and specifically.
- Weight recency over prestige. A senior who hasn't retooled since 2023 is a wasting asset regardless of pedigree; a mid-level who has rebuilt their workflow twice in two years is appreciating.
- Run the work sample in territory slightly outside their stated stack. You're not testing the skill — you're watching the acquisition rate under realistic conditions.
- Beware skill-collector theater: certificates and course lists without shipped artifacts are stock, not slope. The loop must end in something that ran in the real world.
Make learning infrastructure, not a budget line
- 1Build evals first: automated definitions of 'good' for your critical work. Evals turn every experiment with a new tool or model into a measured learning cycle instead of a vibe.
- 2Make AI pairing the default practice, not a permitted exception — engineers, marketers and ops working with agents daily is the retooling; there is no separate course that teaches it.
- 3Reserve structural time for retooling (a day per sprint outlives any 'innovation week'), and protect it the way you protect on-call.
- 4Publish internal proof-of-loop: when someone adopts a better workflow, the write-up plus the reusable config goes in the shared repo. Individual learning becomes team capability.
- 5Retire the annual training budget as your primary lever. Budgets buy courses; infrastructure changes the daily loop — and the daily loop is where half-lives are fought.
