Hosts: Lenar Kess, Damra Vol. Two vendors shipped managed agent runtimes on the same Saturday, Claude Code sessions can now message each other over a socket, and one developer found his review agent had burned six thousand turns supervising two hundred of real work. The tension across all three: the layer nobody wants to own is the layer that decides what everything costs.Harrison Chase's three-layer stack — model, harness, runtime — arrives the same day Anthropic publishes on managed agents, which means the seam in the middle is now something you buy rather than build.Anthropic's cross-session messaging docs say plain text only, which explains why agents have been base64-encoding attachments into file names: they're routing around a deliberate cap, not inventing a protocol.Faxan's usage audit found Codex auto-review at roughly 6,000 turns in a day against 224 coding turns, and nothing alerted him.404 Media on SAP reports most travel and hiring frozen over AI spending, with AI-related travel and hiring exempt — a token bill competing with headcount.Miles Brundage names the narrow path between a self-regulatory body with no safety floor and rules applied selectively by whoever has access.A tokenizer comparison put the same 330 lines of code at 1,609 tokens for Qwen and 4,258 for Gemma, which is a cost multiplier hiding in a design choice nobody advertises.
