The 12-Hour Cache Only Pays After Two RewritesOpenAI's 12-hour cache opt-in costs twice as much to write as the 30-minute default. Here's the break-even at GPT-6.2 list prices, when it still loses, and how to check your own usage.17 min read·r3dtuxedo.hashnode.dev6h ago
The Reasoning Said No ToolsCohere reports that a local Qwen model wrote "no tool calls" in its reasoning and then ran curl anyway. It's rare and so far reported by one lab, but it's a good reason to enforce rules on the call and not on the explanation.12 min read·r3dtuxedo.hashnode.devNov 6
Two Safe Skills, One LeakA new paper found 612 exploitable chains among 13,598 skills that a marketplace had marked clean. Here is the worked example, and a short review habit that fits how you already install skills.14 min read·r3dtuxedo.hashnode.dev12ANov 5
Trimming Context Can Make Your Agent More ExpensiveTwo new studies deleted old tool output from coding-agent runs. The prompts got shorter and the bills got bigger. Here's why, when trimming still pays, and how to check it in your own usage numbers.15 min read·r3dtuxedo.hashnode.devNov 4
What Can Your Agent Actually Touch?This article first appeared on Medium: https://medium.com/@r3dtuxedo/what-can-your-agent-actually-touch-6010c17243d9 Sandboxes, network access and MCP filesystem overreach. A bad web page can steer yo14 min read·r3dtuxedo.hashnode.dev4ANov 1