tagged
#ai
Sub-agents over a persistent context tree Sub-agent context is a tree worth keeping, not scaffolding to throw away. Resume, branch, staleness briefs, and fan-out under one token budget. Your agent's evals should live next to its tools Eval distance is an iteration tax. Why efferent keeps its eval suites in the same workspace as the prompts and tools they judge. Swap your LLM provider at runtime, not compile time Provider choice is request-scoped state, not architecture. How efferent routes one LanguageModel port across Claude, Gemini, and OpenAI per request.