ai-core-runtime

Classify the user intent first, then activate only the minimal agent set needed — for example architect plus validator for design tasks, or debugger plus validator for troubleshooting. Never activate all agents simultaneously, and validate outputs before returning them.

A microkernel runtime keeps the core minimal and treats features as loadable modules. Skills and references are loaded on demand through progressive disclosure rather than preloaded, which keeps context usage low and the system easier to debug.

Use patch-only updates instead of full file rewrites, load references selectively, and compress session state with a continuation-memory approach after roughly twenty turns. Tiered-depth loading fetches only the data depth the current step requires.

Enable extended thinking for architectural trade-offs, complex debugging, and multi-step reasoning chains by setting a thinking budget in the API request. Skip it for simple CRUD operations, formatting, or lookups where the extra tokens add no value.

Skip orchestration for single simple Q&A, one-off code snippets, or tasks that map directly to one specific skill. Routing trivial input through a runtime wastes tokens; respond directly with zero agents activated instead.