Agent Design · Skill Libraries

Your skills work.
Not for the reason you think.

Researchers stopped asking whether agent skill libraries help and started asking why. The mechanism that most builders assume is responsible for the gain accounts for about one in twenty-two helpful uses. The mechanism most builders overlook accounts for roughly two in three. And the retriever that connects libraries to agents falls apart long before content quality can matter.

Core concept
Procedural anchoring: a skill that stabilizes what the agent does step by step provides more lift than a skill that supplies facts the agent didn't have. And retrieval failure is a bottleneck separate from both.

First surfaced in Tandemly Briefing — 2026-08-14.

scroll to explore

Skill libraries help.
Nobody knows how.

Researchers have spent years showing that giving agents access to a library of reusable skills improves task success. That finding is solid. What's missing is the mechanism behind it, and the mechanism determines what you should actually build.

The standard pitch for agent skill libraries goes something like this: store useful knowledge or procedures in a searchable index, retrieve the relevant ones when a task appears, let the agent apply what it finds. Studies confirm this works. Task success goes up when skills are available.

But "does it help?" and "how does it help?" are different questions, and they have very different implications for what to invest in. If skills work by supplying missing facts, you should pour effort into the quality of those facts. If they work by stabilizing action sequences, you should pour effort into the structure of those actions. If they fail because the retriever breaks before the content can even reach the agent, then content quality is largely irrelevant until the retrieval problem is solved.

The field has generated a lot of evidence for the first question and almost none for the second. That gap is what this paper goes after. The researchers describe their goal plainly: not to add another proof that skill libraries are helpful, but to dissect how they are helpful, and to map where that mechanism stops working.

The question this paper asks

When a skill library raises an agent's task success rate, what is actually happening inside the execution trajectory? And when library-equipped agents still fail, is the failure in the content of the skills, the retrieval of the skills, or something else entirely?

8,135 trials.
238 labeled trajectories.

The researchers built a controlled study that held different properties of the skill system fixed while varying others. That structure is what separates the retrieval bottleneck from the content bottleneck, two failure modes that aggregate task-success numbers routinely conflate.

The study normalized 8,135 trial records across conditions that isolated four variables in turn: how skills are represented, how outcomes are annotated, how retrieval difficulty is varied, and how robust results are across different agent harnesses. Varying one while fixing the others is the move that lets you attribute a result to a cause rather than a correlation.

On top of the controlled experiments, the team ran paired trajectory analysis on a sample of cases where skills were helpful and cases where they weren't. Two researchers coded 238 open-labeled instances and resolved disagreements through discussion. The result is a taxonomy of three categories and twelve skill-use modes that describes what is happening inside a trajectory when a skill makes a difference.

One condition in particular is worth noting: the researchers included a setting where the exactly correct skill was provided directly to the agent, bypassing retrieval entirely. This is the control that lets you check whether finding the right skill is sufficient for downstream success. The answer turns out to matter quite a bit for how you interpret offline retrieval benchmarks.

What the taxonomy separates

The three categories distinguish skills that change what the agent knows (knowledge injection), skills that change what the agent does and in what order (procedural anchoring), and cases where the skill was retrieved or invoked but didn't materially affect the trajectory at all. Twelve finer-grained modes sit inside those three categories and describe specific patterns of how each type of influence operates.

The mechanism is procedure.
The bottleneck is retrieval.

Two findings emerged from the taxonomy and the controlled conditions. Both reshape the standard mental model for how to build and maintain a skill library.

Procedural anchoring
65.7%
of helpful skill uses
Knowledge injection
4.5%
of helpful skill uses
Retrieval precision at 100 skills
3.3%
actual-use precision (was 29.6% at 5 skills)
Finding 1: Skills stabilize sequences, not knowledge gaps

Procedural anchoring accounts for 65.7% of the cases where a skill made a meaningful difference. Explicit knowledge injection accounts for 4.5%. The mechanism practitioners tend to assume is responsible for skill library gains is roughly one-fourteenth as common as the mechanism they tend to overlook.

What procedural anchoring looks like in practice: the agent has the relevant information somewhere in its context. What it lacks is a stable sequence of actions to apply that information. The skill provides the sequence. The agent follows it. The task succeeds where it otherwise would have drifted into improvisation or gotten stuck in a loop.

Finding 2: Retrieval fails as the library grows

Actual-use precision collapses from 29.6% to 3.3% as the skill pool grows from 5 to 100 skills. This is not a gradual decline. It is a sharp one. An agent with five skills finds and uses the right one roughly three times in ten. With a hundred skills, it does so roughly three times in a hundred.

The researchers describe this as a failure mode that is separate from the content of the skills. The skills themselves may be excellent. The retriever that locates them degrades under scale. This means teams that are continuously expanding their skill libraries may be unwittingly making their retrieval problem worse while their content problem gets better.

Common assumption
Skills supply missing knowledge. An agent fails because it doesn't have the relevant facts. A well-stocked skill library fixes the knowledge gap, and that's what drives the task-success improvement.
What the data shows
Skills anchor action sequences. Agents most often have the information they need. What fails is the execution path. Skills that specify what to do, in what order, stabilize that path. Knowledge-injecting skills are a small minority of what actually helps.
The offline retrieval proxy problem

The study included a condition where the ground-truth skill was provided directly to the agent without retrieval. Task success in that condition did not reach ceiling. This establishes that exact invocation of the right skill is neither sufficient nor necessary for downstream success.

That result directly undermines the usefulness of offline retrieval benchmarks. Most such benchmarks measure whether a system retrieves the ground-truth skill. They assume that hitting ground truth on retrieval is the right target. The data here suggests otherwise: the agent can use the wrong skill and succeed, and can receive the right skill and fail. Retrieval accuracy on a held-out benchmark is a poor proxy for the thing you actually care about.

Scope and limitations

The taxonomy comes from 238 manually coded instances. Hand-coding at that scale is standard for qualitative trajectory analysis, and two coders resolved disagreements through discussion, but the categories reflect the researchers' interpretive choices. The retrieval-precision collapse is measured within the experimental setup the team designed. Whether it generalizes to other skill formats, retrieval architectures, and task domains would require additional study.

Write procedures,
not encyclopedias.

The mechanism finding and the retrieval finding together point in the same direction: the investment strategy that the field has tacitly adopted for skill libraries may be pointing at the wrong targets.

1
For teams building skill libraries
Write skills as procedural anchors, not fact repositories. Ask of every skill: does this specify a sequence of actions? Does it tell the agent what to do next, in what order, at the decision points where it typically goes wrong? A skill that answers yes to those questions is likely to provide lift. A skill that mostly encodes facts the agent could have inferred from context is unlikely to.
2
Track actual-use precision as the library grows
Aggregate task success is not the right metric for a skill library's health. Log which skills are retrieved at query time, which are actually used in the execution trajectory, and whether the task succeeded. The gap between retrieved and used is your retrieval signal. If actual-use precision is falling while task success is holding steady, you may be getting lucky on some tasks while the retrieval problem accumulates.
3
Stop trusting offline retrieval benchmarks as a proxy for downstream success
The paper establishes that exact ground-truth retrieval is neither sufficient nor necessary. An offline benchmark that measures whether the system finds the right skill is measuring the wrong thing. The right target is actual-use precision on real task trajectories: did the retrieved skill materially influence the execution path, and did that influence help?
4
For AI researchers and tool builders
The retrieval-precision collapse at 100 skills is a problem that requires a targeted solution. Better embedding models or re-ranking layers may help, but the pool-size scaling suggests the issue is structural. The Skill-Is-Not-Document work on set-level compatibility (queued separately) may be a relevant parallel: retrieval for skill sets may need to optimize for collaboration across retrieved skills, not just per-skill relevance.

Where to go
from here.

The paper is a diagnostic, not a design prescription. Here is where to start applying what it found.

1
Audit your existing skills for the procedural anchoring pattern
Read through your skill library and sort each entry: does it specify a sequence of actions at decision points where the agent typically struggles, or does it mostly inject facts? The latter category is unlikely to produce consistent gains and may be worth redesigning before expanding the library further.
2
Run the pool-size precision test
Fix a held-out task set. Run your agent with skill pools of 5, 10, 25, 50, and 100 skills. At each pool size, measure actual-use precision (how often a retrieved skill appears in the execution trajectory and influenced it) alongside task success. This generates the scaling curve the paper produced and tells you where your retrieval starts failing.
3
Build a trajectory annotation pass
Pair the pool-size test with a lightweight trajectory review. For each failed task where a skill was retrieved, ask: was the skill used? Was it the right skill? Did using it help? Even a sample of 50-100 trajectories hand-coded with those three questions will tell you whether your failure is in the content, the retrieval, or the application of what was retrieved.
4
Read the paper
Jiang, Z., Huang, F., Xing, H., Wang, M., Liu, S., Li, Y. et al. (2026). Demystifying Agent Skills: Why They Work-Until They Don't. arXiv:2608.14036. The taxonomy of three categories and twelve skill-use modes is the main deliverable; it gives precise vocabulary for talking about which part of a skill system is producing the failure you're observing.
5
Related work in this space
For the skill lifecycle: MUSE-Autoskill (gating admission with unit tests, retiring underperformers). For set-level retrieval: Skill-Is-Not-Document (compatibility-aware two-stage routing). For toolset composition: Atomic Actions to SOPs (promoting recurring sequences into named composite tools). This paper sits above all three: it diagnoses why the retrieval and content layers interact the way they do before any of those interventions are applied.