lazy-agent-loader

Load agent summaries first and fetch full definitions only for selected agents. This lazy-loading approach keeps context lean during agent selection, loading complete agent data on-demand to minimize token consumption in agent workflows.

Use a summary-first index paired with on-demand loading of full definitions. Load agent metadata initially, then retrieve complete definitions only when an agent is chosen as primary, reducing unnecessary context overhead.

Yes. Session-level caching stores loaded agent definitions in memory, eliminating duplicate fetches within a single session. A force-reload capability lets you refresh cached agents when needed.

Store full agent definitions in separate files using the pattern agents/[agent-id].md. A metadata index points to these files, enabling the loader to fetch definitions only when a specific agent is selected.

Use on-demand loading when you have many agents but only evaluate a subset during each workflow run. This threshold-based approach saves tokens by deferring full-definition retrieval until an agent becomes a candidate.