Research
Emergent role classification from skill mapping
Every few years a new engineering role forms where two older ones overlap. Dev plus Sec plus Ops became DevSecOps. A software engineer who picked up statistics and model training became an ML Engineer. This study detects these new roles as they form by analyzing 47,101 Fortune 500 software job postings against a map of what each established role was historically supposed to require.
What we set out to find
Job titles lag the work. A company hires a "data scientist" and then asks them to ship LLM agents; it posts for a "security engineer" who turns out to live inside the deployment pipeline. The skills move first, and the title catches up years later. At Andela we're obsessed with identifying these shifts early and this study is a lens into how we catch that movement while it is happening.
This is how the ML Engineer formed in the gap between software engineering and statistics, how DevSecOps formed between development, security, and operations, and it may be how the AI Agent Developer is forming right now between backend engineering and LLM orchestration (spoiler: our data says not yet; see finding 02). A role emerges in the seam between two older ones as their skill sets bleed together.
Detecting emergence without a clock
Emergence usually means change over time, but this corpus is a single five-week snapshot (30 March to 2 May 2026) of job postings pulled from current, publicly facing Fortune 500 job boards. So we reframed "historical" as structure rather than date with every established role requiring a certain set of skills. These skill mappings are from Andela's internal skill taxonomy, which we use to identify emerging skill bundles. We measure how far the postings have drifted from that baseline, and we date the technologies involved: the recency score rewards bundles whose tools were born in the last three years (2023 or later, against a corpus collected in 2026) and whose median technology is young, while a bundle built on decade-old tools is a hybrid that already settled. None of this run's bundles has a post-2023 median, so "emerging" on this page means a median technology birth year of 2018 or later. Canonical drift plus technology vintage stand in for the missing time axis.
We have a wider thesis on talent debt, which is that skill half-lives are shrinking. We cannot measure half-life on a five-week window, so it is not a focus for this research, but we can see which skill bundles are fusing into new roles, which is the leading edge of the same phenomenon.
What the data shows
The strongest signal is a role still settling into its name
The MLOps Pipeline Engineer, who builds and runs the automated infrastructure to deploy, version, and monitor machine-learning models in production, scores highest on emergence. It bridges five established roles, principally ML Engineering, DevOps, and Data Engineering, pulling together MLOps, CI/CD for ML, model registries and versioning, workflow orchestration, and ML observability. Its skill bundle appears in 1,582 postings (postings carrying at least two of the bundle's skills from different established roles, not postings titled for it), and the validation below matches it to MLOps Engineer, the youngest of the known emergent roles it recovered: the detector's top candidate is a role caught while its title is still consolidating across enterprises, not a brand-new invention.
The AI frontier is narrower than the hype
Exactly one foundation-model bundle ranks among the most-emergent candidates: the LLM Application Engineer, the contemporary AI engineer building on foundation models rather than training them, third by recency across 6,758 postings that carry part of its bundle. Once every bundle's technologies are dated, no candidate's stack has a median birth year later than 2019: Fortune 500 hiring trails the technology frontier by several years, and the newest tools (post-2023 agent frameworks and the like) have not yet stabilized into any recurring cross-role bundle in this corpus.
Employers name the old role, hire the new one
Among the 1,832 postings whose title names one of this bundle's home roles (chiefly "AI Engineer" and "ML Engineer"), 53% (972) already carry at least two of the bundle's skills from different home roles, the same cross-role test used for coverage, with large language models, conversational AI, the OpenAI API and Vertex AI, and evaluation harnesses such as OpenAI Evals among the most common bridging skills. The title names a role employers already know; the skills are drifting toward one that has no settled name yet.
How we detected emergence in a snapshot
Through seven steps we turn a pile of job postings into a ranked list of forming roles.
Baselines
Build each role's "before" two complementary ways: taxonomy skill centroids (external) and the profiles of title-pure postings (empirical).
Home roles
Assign every skill the established role it historically belongs to, keeping only confident placements.
Bleed metric
Keep skill pairs that co-occur far more than chance, span different home roles, and are not synonyms.
Communities
Cluster the bleed graph alone, so every community bridges two or more historical roles by construction.
Vintage
Date each bundle's technologies; recent stacks rank as live emergence, old stacks as settled hybrids.
Title divergence
Measure how often a posting's title names one role while its skills show the cross-boundary bundle.
Label
A local language model names each isolated bundle from its skills and an example posting.
Why a snapshot can detect emergence
In an earlier attempt, I recombined clusters drawn from the same co-occurrence graph and simply kept re-finding the obvious. The method presented here breaks that loop by anchoring the "before" to a role taxonomy and to title-pure postings that are independent of the Fortune 500 co-occurrence patterns we are searching. Bleed is then the measurable gap between what a role was supposed to require and what employers are actually asking for.
What is distinctive here
- Cross-sectional bleed. Emergence is recovered from canonical-versus-observed drift in a single snapshot, not from posting-volume growth over time.
- Technology-vintage scoring. Dating the stack separates genuinely emerging bleed from long-settled hybrids; the vintage axis stands in for the missing time axis. In the future we will incorporate temporal signals directly into the model (if we can get the data).
- An embedding-distance synonym filter. Requiring the two bleeding skills to be semantically distant removes the near-synonym noise that defeats volume-based methods.
Hybrid-job research already exists (Lightcast, O*NET, ESCO). We do not claim to have invented hybrid roles; the contribution is a snapshot-compatible, taxonomy-anchored, vintage-scored detector.
The candidate landscape
Twenty-three candidate roles, read along two independent axes. Bleed asks how far a bundle reaches across established roles, irrespective of age. Recency asks whether a bundle is built on new technology, which is the primary ranking, and where the newest stacks rise to the top. The two rarely coincide: the newest bundles aren't the most cross-cutting, and the bundles that bleed the widest are often long-settled. A low recency score isn't a weak finding but a real cross-role bundle that emerged years ago and stuck.
The bleed graph
Each point is a skill. A line connects two skills only when they co-occur far more than chance, belong to different established roles, and are not synonyms, which is the signal of skill bleed. Communities in this graph are the candidate roles: each dense cluster is one bundle. Shown here: a 200-skill subsample carrying 2,713 of the full graph's 4,093 edges. Each skill is coloured by the candidate bundle it belongs to; grey skills survive the bleed filter but fall outside any kept community. You can drag, scroll to zoom, and hover any point.
A closer look at the top 10
The ten candidates ranked highest on recency, from the newest to the freshly settled. Ten is where the scores break: the tenth still sits at 0.111, the eleventh drops to 0.044. This is deliberately a different ten from the widest-bleeding bundles charted just above: recency asks how new a bundle's tech is, bleed asks how far it reaches across roles, and the two rarely pick the same leaders. Four of the ten (GitOps Platform Security Engineer, Governed BI Engineer, GitOps Delivery Security Engineer, AWS Data Platform Engineer) were judged long-settled hybrids in the July audit; they are here because their tooling is recent, not because the role is new. Each card shows the established roles the bundle draws from and the skills doing the bridging.
Every candidate, ranked
All 23 detected bundles, sorted by Recency (how new the tech is) by default, the page's primary ranking; the # column is that recency rank, the same number the findings and validation sections refer to. Click any column heading to re-sort, including by Bleed (how far each crosses established roles). Each recency cell shows how many of the bundle's top skills carry a datable vintage; a median resting on one or two dated skills is a weak estimate. Coverage counts postings carrying at least two of a bundle's skills; bundles overlap, so coverage summed across all 23 exceeds the 47,101-posting corpus. "In canon" marks bundles whose name resolves to a job title already in common use (Security Engineer, Cloud Engineer, Analytics Engineer…); "in the wild" marks bundles no established title fits, so the name is descriptive. This is a check against a short list of established titles, not a lookup in O*NET, ESCO or Lightcast. Two candidates (MLOps Pipeline Engineer, Lakehouse Analytics Engineer) were marked in canon by hand because MLOps Engineer and Analytics Engineer are established titles despite the generated name. One candidate (Android XR Developer) is flagged as a detection artifact from the manual audit; see the limitations section.
| # | Emergent role | Recency ▾ | Bleed | Postings |
|---|
How we know the signal is real
The tests below were specified before the candidate list was inspected, with two post-hoc refinements made and disclosed: the recall signatures (including a rename-only realignment after the taxonomy v1.3.0 migration) and the clustering resolution (final values in the reproducibility note at the end of the page).
Behind the bleed metric, every skill's home-role assignment had to survive the two-baseline check: of the 2,026 scored skills, 596 were placed identically by both baselines, 432 agreed within the top three, 393 rest on the taxonomy alone (no title-pure evidence either way), and 605 low-confidence placements were excluded from the graph entirely.
The seven roles we knew should appear, and five did
What this study does not claim
It is a snapshot, not a trend
Emergence is inferred from canonical drift and technology vintage, never from posting-date growth. The corpus spans about five weeks.
It reflects Fortune 500 employers' hiring
The corpus is large-enterprise and software-role filtered, and postings were selected by employer (Fortune-ranked companies), not by location, so a share are for roles outside the United States. Findings generalize to that population, not the whole labor market.
Skill extraction is imperfect
An importance floor, a canonical-name map, and a blocklist reduce extraction noise but do not eliminate it; residual false positives (e.g. design tools inside the LLM bundle) stay visible in the skill lists rather than being silently removed. A single-rater manual audit of all 23 candidates (July 2026) flagged one as an artifact: Android XR Developer, whose cross-home edges come from the taxonomy splitting Android tooling across Mobile and VR/AR home roles. It is kept visible and flagged in the catalog rather than hidden.
Technology dating is fuzzy
Birth years are judgment calls drawn from a versioned, auditable lookup and the emergence score is a rank, not a precise date. Several bundles' medians rest on only one or two dated skills; the catalog shows each bundle's dated-skill coverage.
Ubiquitous skills carry no signal
A skill that co-occurs with nearly everything produces no bleed edge by design, so it is absent from the graph. A feature of the method, not a gap.
Hybrid jobs are prior art
The phenomenon is well established in the literature. Our contribution is the snapshot-compatible, taxonomy-anchored, vintage-scored detection method.