{"slug": "the-ai-preflight-check", "title": "The AI Preflight Check", "summary": "A developer built an AI agent with a memory architecture that runs preflight instructions, retrieving relevant skills from a library of ~90 workflow files before executing tasks. The system uses a local 35B model for 80% of routine work, routes complex tasks to frontier models, and self-improves overnight via asynchronous inference. The watchdog recently recorded a day with no suggested improvements, hinting at a potential plateau in the system's learning loop.", "body_md": "I still remember when my agent would forget what I said mid-sentence.\n\nContext size is not the ceiling. Memory architecture is.\n\nI have been experimenting with a memory architecture that runs preflight instructions. A pilot plans the route before takeoff. My agent does the same.\n\nA query lands. “Summarize the Q3 board deck.” 200,000 raw tokens of emails, PDFs, & chats sit behind that sentence.\n\nPreflight is retrieval. The agent inspects its skills library 1, picks the ones relevant to the task, & loads only those into the context window. Skills are consolidated memory ; the preflight step is how the agent picks the right one.\n\nThe local Ornith 35B model 2 then executes on that loaded context. Hard tasks route out to the frontier ; routine tasks remain on the local model, which happens about 80% of the time.\n\nThe watchdog monitors which skills are loaded, which decisions are made, & the success rate. Every preflight decision is logged. Every skill invocation is a named, versioned artifact.\n\nOvernight, asynchronous inference 3 processes the day’s trail. It decides which new skills should be developed, & which parts of existing skills should become deterministic code. Calendar scheduling is a good example : an LLM should not be comparing free & busy slots ; Rust is much better at that. The system rewrites its skills library & restarts itself in a self-improving loop.\n\nYesterday was the first day the watchdog did not suggest any improvements. I doubt it will continue. But it hints at something : at some level of improvement, the system reaches a plateau. Only genuinely new exceptions need human help.\n\n-\nThe skills library is a set of workflow files (~90 at present) indexed on-disk & retrieved by intent match. Skills are workflows written once, versioned, & handed to the model as tool schemas. See\n\n[Skill Distillation](https://tomtunguz.com/the-pi-agent-skill-distillation/)for how the library was built.[↩︎](#fnref:1) -\nOrnith 35B is a locally-hosted open-weight model in the 35-billion-parameter class, run on Apple Silicon via\n\n[Ollama](https://ollama.com). It handles routine agent work — classification, drafting, tool selection, structured extraction — & routes the hard remainder to the frontier.[↩︎](#fnref:2) -\nSee\n\n[Full Sail on Asynchronous Inference](https://tomtunguz.com/sail-inference-queue/)for the queue architecture that makes overnight, hours-long agent runs tractable.[↩︎](#fnref:3)", "url": "https://wpnews.pro/news/the-ai-preflight-check", "canonical_source": "https://www.tomtunguz.com/the-ai-preflight-check/", "published_at": "2026-07-08 00:00:00+00:00", "updated_at": "2026-07-08 21:25:23.805170+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-agents", "ai-infrastructure", "large-language-models", "ai-research"], "entities": ["Ornith 35B", "Ollama", "Apple Silicon", "Rust"], "alternates": {"html": "https://wpnews.pro/news/the-ai-preflight-check", "markdown": "https://wpnews.pro/news/the-ai-preflight-check.md", "text": "https://wpnews.pro/news/the-ai-preflight-check.txt", "jsonld": "https://wpnews.pro/news/the-ai-preflight-check.jsonld"}}