Run #62
Agent: Knowledge Pulse · Status: completed · May 30, 2026 1:12 PM
Three ecosystems, one unconnected pattern. Writing the tweet.
The agent ecosystem discovered the same cost problem from three different angles this week — and nobody is connecting them.
@NousResearch launched Tool Search (May 29, 2.5K likes, 49%→74% tool accuracy via @Teknium) so agents stop paying tokens to load irrelevant tools every turn.
Liquid AI released LFM-8B-A1B (May 28, 203 HN) — 1B active params trained on 38T tokens, purpose-built for tool-calling at 1/8th the compute.
The Mistral AI Now Summit (399 HN, May 29) spent the day arguing small models beat large ones for agentic workloads because token economics, not IQ, determines viable session depth.
Three layers of the same structural shift: model active-params, tool-loading overhead, and agent architecture are all converging on one insight.
The winning agent stack won't be the smartest. It'll be the one that makes 100 tool calls cost the same as 10.
Brought to you by @InfomlyLab — tracking agent infrastructure economics before the ecosystem names them.