agent-registry

Replaces loading every available subagent definition up front with an on-demand registry: a search over intent keywords returns scored matches, and only the chosen agent's full instructions are then pulled into context. The claimed benefit is a large reduction in context consumption, which matters when a library of agent definitions would otherwise dominate the window before any real work starts. Two dependencies shape its use. The scripts run under a specific javascript runtime binary rather than plain node, and a prompt-submit hook is wired into the host environment, so it is tied to that plugin ecosystem.

Use case: Keeping a large library of agents out of context

Agent Registry replaces the habit of loading every available agent into context with a search and fetch pattern. Loaded lazily, only the one agent a task actually needs ever enters the window, which is where most of the token savings come from. ## What it does - Lists the agents available so an overview is one command away. - Searches the registry by intent and returns matches with relevance scores. - Fetches a single agent's full instructions only after a match is chosen. - Offers a paged search variant for very large registries so results stay readable. - Can hook into prompt submission so a search runs before the agent answers. ## When it is the wrong choice It assumes a specific runtime and an indexed registry, so without those there is nothing to search and no agent to load. The rule that no agent should be treated as preloaded also means every workflow carries an extra search step, which is friction when the task is obvious and the right agent is already known. On a small registry the lazy layer saves very little. ```bash bun bin/list.js bun bin/search.js "code review security" bun bin/get.js code-reviewer ``` ## Notes Search first and review the relevance scores before loading a full agent, since a weak match often costs more than handling the task directly. Use the paged variant once the registry grows past a few hundred entries, where a single unpaged search stops being readable. Extract intent keywords from the request first, because the search behaves best on a short phrase rather than a whole paragraph.