Token Counts are Obsolete for Agentic Workflows
Agentic loops change the economics of inference.
Agentic loops change the economics of inference.
The paper LLM-as-a-Judge Is Not an Oracle signals a shift away from model self-evaluation.
The paper On the Inseparability of Instructions and Data in Shared-Embedding Sequence Models proves prompt injection is a fundamental flaw.
The week infrastructure constraints collided with AI agent ambitions. Builders are hitting the memory wall, paying tool-calling overhead, and rediscovering that reliability is engineering, not prompting.
Three iterations of a substrate benchmark forced three rewrites of what agent-native CLI means. Operation size and model capability dominate. Substrate format is a few-percent margin call.
Smaller models can now carry real agent workloads. AgentFloor gives you a measurable way to decide which pipeline steps need a frontier model and which run on an 8B.