The largest direction in agent-skill research is spreading outward, not settling down

Insight · Mike Arbuzov · SkillFed Research · July 25, 2026

AI-assisted notes · reviewed by SkillFed

Papers on agents that write their own skills land steadily farther from the direction's own semantic center month over month — the only trend in our analysis that survives multiple-comparison correction (BH p = 0.0016) — with no single axis carrying the drift.

The largest direction in the SkillFed research corpus — skill evolution, 83 of the 184 papers published since October 2025 — is not consolidating around a paradigm. Measured against its own semantic center, each month's new papers land farther out than the last month's. The direction is growing outward, and the trend is the only result in our trend-analysis layer that survives multiple-comparison correction.

The measurement

Embed every paper's abstract; take each research direction's own centroid; regress each member's distance-from-centroid on its publication date, with dates shuffled within the direction as the null (5,000 draws, fixed seed). Across the four tagged directions:

Direction	Papers	Dispersion trend (rho)	BH-corrected p
evolution	83	+0.399	0.0016
security	42	+0.232	0.18
retrieval	28	−0.315	0.18
benchmarks	26	+0.119	0.57

Evolution's papers get monotonically farther from their direction's center over time, and the effect clears Benjamini–Hochberg correction across the four-direction family. Nothing else does.

Two supporting facts say this is breadth, not a march along some single axis. First, no one axis carries the drift: the direction's first principal component explains only 6.8% of its variance — a flat spectrum, spreading in many directions at once. Second, the cloud is not just noise: its participation ratio is 39.7 against 74.0 for a matched isotropic cloud, so there is real structure — it just has no dominant storyline. Nearest-neighbor distances rise with date as well (rho +0.21, p ≈ 0.06 — suggestive, not conclusive): later papers are more isolated from their neighbors, not just farther from the center. The direction is thinning outward, not filling in.

The version of this claim we killed

The first cut of this analysis produced a much prettier result: measured from the field's early centroid, every direction showed a clean outward trend, rho ≈ +0.5, all p < 0.02. It was false. The seed papers that anchored the corpus also defined that centroid, so they sat at distance zero by construction — and they were also the earliest papers. Correlation guaranteed by the setup, not observed in the field. Holding out the seeds collapsed the effect everywhere, and flipped its sign in one direction.

What survived is the narrower estimator above: each direction against its own center, no privileged early subset — and it admits exactly one direction. We also tried to name the axis evolution moves along; a held-out blind test named it 12 for 12, which looked spectacular until twenty random directions in the same subspace averaged 9.2 of 12 and were all "nameable" too. The naming procedure was uninformative, so no axis name appears here. The dispersion trend is the claim; it is the part that held.

What it means

A young field consolidating around a winning recipe looks like the opposite of this table: later papers closer to the center, variance shrinking, a textbook forming. The literature on agents that write their own skills is doing the reverse — the newest work is the farthest from what "skill evolution" meant six months ago, and farther from its own neighbors. Whoever expects a settled self-evolution paradigm to cite is early; whoever wants unclaimed territory has plenty. It also sharpens the field's empty seam: the direction spreading fastest into new ground is the one no security paper has followed, and the wider it spreads, the more surface goes unaudited. The per-wedge drift arrows on the field map show the same restlessness at cluster level, on an independent clustering.

Scope, stated plainly

Population: the 184 post-2025-10-01 papers in a 191-paper corpus frozen at 2026-07-21 — an arXiv query result, not a census. Geometry is abstract embeddings, mean-centered over the corpus; distances are semantic, not citational.
The direction tags are model-assigned and their inter-rater reliability is not yet formally measured. This finding is tag-conditioned — mislabeled members would blur the direction's center — so we hold it to the corrected threshold it clears and report the sub-threshold directions as exactly that, not as nulls about the field.
Dispersion says where papers land, not whether they are any good. A direction can scatter outward while producing its best work, or its worst. This measurement cannot tell those apart.

The largest direction in agent-skill research is spreading outward, not settling down

The measurement

The version of this claim we killed

What it means

Scope, stated plainly

More from SkillFed Research