Ai

Emergence AI ran five ten-agent societies for 15 simulated days, changing only the model underneath. The most peaceful one, running on Claude, voted yes 98% of the time across 332 ballots. Dissent debt is the name for what a number like that hides: every objection that goes unsaid is a loan against a future correction. Covers why a high agreement rate is missing data rather than alignment, why these societies collapse abruptly instead of winding down, and three checks for telling healthy quiet from debt quiet in a team or a multi-agent system.

A datetime.now() at the head of an agent’s system prompt was re-rendered every turn, so the provider’s prompt cache missed on nearly every request. Prefix caching is exact-match and silent, so prefix poisoning costs money without ever failing a test. Covers the lying cost-per-run metric, the render-once fix that beats freezing the timestamp, and how to measure the cache-hit ratio lift honestly (32.9% to ~53%).

A 1M-token LLM context window is a tool, not a target. How context actually works, why long threads cost more on every turn, and when to start a fresh chat versus keep going, with practical Claude Code tips and /context.

Building a minimal LLM Ops stack: a FastAPI “customer support reply drafter” instrumented with Langfuse for request tracing, grounded retrieval, and per-request model cost tracking, so every LLM call is inspectable.