🎙️ State of Agentic Coding with Ben Vinegar and me is out. - AI writing and watermarking - Cost of inference and economics - What's up GitHub? - Agents can get anything to work?

Also on Youtube: And Spotify: RSS is on the website:

@bentlegen Could we get you guys on Apple Podcasts? 🙏

@bentlegen best video of the week (im going to view it now)

@bentlegen inference costs will dictate how agents actually scale

@bentlegen On the economics bit, the number that surprised me is that agent cost tracks retries rather than price per token. A model that's 20% cheaper but needs a second attempt on a third of tasks ends up dearer, and you only see it once you log attempts per finished task.

@bentlegen If agents can get anything to work, the work doesn't disappear. It moves. When building was hard you only tried two or three things. Make it cheap and you try twenty. Now you're picking between prototypes instead of coding them.

@bentlegen Curious how the economics of inference shift when agents start writing more than reading code.

@bentlegen Are the agentic coding costs mostly from repeated inference calls, or the context windows getting really long?

@bentlegen AI 写作 watermark 让我更在意“产出归因”而不是模型能力。工程团队如果无法把 agent 生成的 diff、上下文和测试结果留档，后面无论是安全审计还是成本核算都会变成猜；GitHub 的下一步其实就是把这些 provenance 变成一等数据。

@bentlegen nice! I never miss these. Thanks for making more episodes :)

@bentlegen “Agents can get anything to work” is the demo frame. The industrial question is what they’re allowed to send, charge, or write when the blast radius is a real company — and how fast the fleet can move once those gates hold.

@bentlegen super curious to see how the economics of inference will shift as agents start writing more code than just reading 👀✨