0:00 UC Berkeley and Microsoft researchers just proved an external controller can slash coding-agent token costs by thirty-one percent across sequential codebase tasks.0:11 In episode two thirty-six we tracked DeepSeek's benchmark scoring splits; Today, the timeline shifted from raw model bragging rights to the plumbing that keeps agents from bleeding cash.0:24 Look, while everyone else argued about model size, the real signal was moving into infrastructure and risk.0:32 Censys flagged active exploitation of a critical Atlassian Confluence flaw—rated nine point eight out of ten—monitoring one point two million exposed properties, though exposure alone doesn't prove compromise.0:47 Meanwhile, the U.S.0:48 Commerce Department dropped an interim rule choking off remote access to advanced AI chips through third-country data centers.0:57 Over in enterprise tooling, Perplexity committed ten thousand GPUs to a dedicated Frontier Research lab aimed at building an automated researcher, and ElevenLabs absorbed Kotoba Tech, Japan's leading voice AI lab.1:12 Even OpenAI agreed to let evaluation nonprofit METR run an independent alignment probe on its largest reinforcement learning run to date.1:22 Now, here's the thread you actually need to pay attention to.1:26 Developer Thariq dropped a post arguing that software factories are merely a temporary problem.1:33 His point?1:34 Models will learn to construct those pipelines themselves, meaning the real bottleneck isn't building automated factories—it's designing workspaces where humans and agents can actually collaborate.1:48 Almost immediately, multi-agent framework Swarms launched AgentRearrange, proving his point by compressing entire agent workflows into a single line of code with just two operators.2:01 So what does it all add up to?2:03 Well, here's the thing.2:05 If models build their own orchestration, the constraint shifts entirely to cost and verification.2:12 That's why that Berkeley paper, TokenEconomizer, hit like a bomb on developer feeds.2:18 Right now, coding agents resend their entire context history on every single turn.2:23 Run an agent across an entire repository, and your bills explode.2:28 TokenEconomizer acts as a black box sitting outside Claude Code, Codex, and GitHub Copilot.2:35 It learns which unused tools to prune, which repo facts to inject, and when to kill a failing loop.2:42 Across two hundred seventy-seven SWE-Bench tasks, it lopped eighteen to thirty-one percent off average token spend—and improved Pass at One accuracy in four out of five settings.2:54 Wait.2:55 Think about that math.2:56 We spent two years treating agent overhead as an inevitable tax.3:01 And so — and so the real pivot this week is that builders stopped accepting brute-force compute as the answer.3:09 You don't need a smarter model to fix agent bloat; you need a strict referee standing outside it.3:15 Next quarter won't turn on who trains the biggest frontier brain, but on who keeps autonomous agents from bankrupting the teams running them.