JevAgent preflight · planned
JevAgent Skill Preloading
This page offers a clear overview of JevAgent Skill Preloading. It helps you explore this topic and find a useful next step.
1. This capability is still in development
Skill preloading is a planned JevAgent capability, not a feature available in every released SDK. Check the SDK source and release notes before relying on it. This page describes the intended workflow, not a released configuration API.
The proposal places skill candidates within JevAgent's alignment settings. The exact public setting and supported skill sources must be confirmed from merged SDK code before use; no runnable configuration is documented here.
2. Evaluate candidates, then load the useful instructions
The application supplies candidate skills. For the current request, Jev asks a focused usefulness question about each candidate. The runtime then loads the full instructions for the candidates that pass and makes those instructions available before the main agent begins the task.
The check is per skill: Jev should judge whether that skill's described guidance can help with this request, rather than being asked to solve the task or to reason across a large bundle of unrelated skills.
The application makes candidate skills available to JevAgent.
Before the main agent run, Jev evaluates each candidate against the user's request.
The runtime loads only the instructions selected as useful for this request.
The main agent starts with the request and the selected skill instructions.
3. Keep candidate instructions out until they are relevant
Loading every candidate skill's complete instructions into every request can spend context on guidance that does not apply. Preloading is intended to let developers attach a wider candidate set while adding only the skill instructions Jev selects for the current task.
That is intended to reduce avoidable context-token use and preserve room in the context window for task material and selected guidance. This is a design goal, not a measured result, and it does not guarantee a fixed token reduction, better accuracy, or lower total cost: the result depends on the number and size of candidates, the request, and the decisions made for that run.
4. Make each relevance check easy to answer
Jev is most useful when a question gives it the state and definitions needed to recognize one property. The SDK's asking-jev-questions guidance applies that discipline: explain what the request and skill description mean, ask whether the named skill's guidance fits the request, and keep unrelated judgments out of the question.
The selection result is a relevance signal, not proof that the skill is correct, complete, or safe for the task. The main agent still needs to follow the selected instructions alongside its system prompt, permissions, and ordinary task constraints.
5. Skill-source adapters are separate work
The preflight-selection logic and the adapters that retrieve skills from outside the SDK are separate implementation steps. GitHub URLs, Anthropic's Skills API, skills.sh, and other remote catalogs are not supported by this documentation preview; their availability depends on separate adapter work and the SDK release that includes it.
Until the feature and a source adapter are merged and released, do not treat the proposal as an available import workflow. This guide intentionally omits a code sample because the public settings and provider interfaces are not yet verified.