Byung-Gon (Gon) Chun, founder and CEO of FriendliAI, argued in his Day-3 talk that agentic inference is structurally different from chat inference and requires rebuilding the inference cloud around it. FriendliAI, born from a Seoul National University research team that invented continuous batching and whose Orca work inspired vLLM, positions itself as "the frontier AI inference cloud for agents," headquartered in San Francisco with a team in Seoul. Chun claimed 2026 is the year agents go into massive production, driven by explosive agent adoption and open-weight models reaching frontier quality, requiring optimization of end-to-end task latency rather than single-request latency. On an identical coding-agent task building a tower-defense game, GLM 5.2 on FriendliAI cost 27 cents versus roughly $150 for Anthropic's Opus 4.8 — about 5.6x cheaper — while both produced output "clearly usable for many agentic workflows." The talk covered four engineering pillars — prefix caching, KV cache management, cache-aware routing, and agent-aware optimization — with Kilo Code claiming FriendliAI was consistently 7x faster with significantly lower error rate in a split test of GLM-5 usage. Chun argued the unit of work changed from a request to a task, which may involve many model calls and tool calls and run autonomously, so users care when the whole task completes.