Your AI Workflow Shouldn’t Depend on Your Memory (AI & Oats #44)

Running several AI agents at once creates a simple problem: you cannot reliably remember every thought that appears while another task is underway. In this episode, I explain why I prefer to agree a proper plan before significant work begins, then let Claude Code or Codex run. If another thought occurs, I send it immediately as a queued message, or interrupt the task when it changes the direction. I also look at whether a physical desk display for agent status would be useful, when Full Access is worth the trade-off, and why stale context can become a bigger problem than forgetting. I start with some first impressions of GPT-6 Astra, including the website work I have tried and my dad's early verdict on its maths ability. SEE HOW AI COULD WORK FOR YOUR BUSINESS Work with me to find out where AI could help your business, and what to do next: https://duxai.co.uk/contact JOIN ALISTAIR'S AI OPERATORS Step-by-step lessons on Claude Code, Codex and practical AI systems: https://www.skool.com/alistairs-ai-op... CHAPTERS 00:00 Intro 01:01 First impressions of GPT-6 Astra 01:50 Do you need an agent-status display? 04:19 Stop relying on memory 06:10 Full Access and approval prompts 08:37 When agent context goes stale 10:28 Outro #AI #Codex #ClaudeCode