The AI Preflight Check

2026-07-08 · Source: Tomasz Tunguz · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Software Development & Engineering · Depth: Advanced, quick

Summary

An AI memory architecture, dubbed the "AI Preflight Check," enhances agent performance by dynamically managing context. This system employs a preflight step to retrieve relevant "skills" from a library of approximately 90 workflow files, loading only necessary information into the context window for a local Ornith 35B model. This 35-billion-parameter open-weight model, running on Apple Silicon via Ollama, handles about 80% of routine tasks, while complex queries are routed to a frontier model. A watchdog component continuously monitors skill usage and decisions, logging every action. Overnight, asynchronous inference processes this data to identify new skill development needs or convert skill components into deterministic code, enabling the system to self-improve by rewriting its skills library. The system recently achieved a day with no suggested improvements, indicating a potential plateau in its self-optimization.

Key takeaway

For AI Architects designing robust agent systems, prioritize dynamic memory architectures over merely increasing context window size. Implement a "preflight check" system that retrieves specific skills from a library, routing routine tasks to efficient local models like Ornith 35B. Establish a watchdog and asynchronous inference loop to continuously refine skills and convert deterministic logic into code, fostering self-improving agents.

Key insights

Dynamic memory architecture with skill-based retrieval and self-improvement surpasses raw context size for AI agent performance.

Principles

• Memory architecture dictates agent performance.
• Self-improvement through skill monitoring.
• Route complex tasks to frontier models.

Method

A query triggers preflight retrieval from a skills library, loading relevant context for a local Ornith 35B model. A watchdog monitors execution, feeding asynchronous inference to rewrite and improve skills overnight.

In practice

• Implement a dynamic skills library.
• Log all agent preflight decisions.
• Convert LLM logic to deterministic code.

Topics

• AI Agents
• Memory Architecture
• Skill Retrieval
• Ornith 35B
• Self-improving Systems
• Local LLMs
• Asynchronous Inference

Best for: AI Scientist, Research Scientist, AI Engineer, Machine Learning Engineer, AI Architect

Related on AIssential

• A Practical Guide to Memory for Autonomous LLM Agents — Memory architecture, particularly its management, is more critical to LLM agent performance than model choice.
• The Next AI Bottleneck Isn’t the Model: It’s the Inference System — Misattributing AI system failures to the model rather than inference architecture is a common, costly mistake.
• I Thought One Prompt Could Do Everything… I Was Wrong — Multi-agent AI systems, mimicking human teamwork, overcome single-model "context overload" by distributing tasks.
• ReasoningBank: Enabling agents to learn from experience — ReasoningBank enables AI agents to learn generalizable strategies from both successes and failures for continuous self-evolution.
• Self-improving Memory for Agents - Perplexity — Brain enables AI agents to self-improve by learning from their past actions, successes, and failures, enhancing future task performance.
• Week Ending 4.12.2026 — AI progress demands better benchmarks and frameworks for real-world reliability, judgment, and continuous learning.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Tomasz Tunguz.