This week Europe get's it's own office suite just when AI might be making office suites irrelevant, AI+Quantum might be quite good, and we get a new euphemism: Lower Value Human Capital.
On to this week's main stories:
- GitHub Slashes Agent Workflow Token Spend up to 62% with Daily Audits and MCP Pruning. Almost every team I know that's extensively using AI for coding is dancing around token limits. It seems that even true inside the hyperscalers and probably the AI labs (any tokens they use can't be sold to end users). The current plans, API costs, and refresh windows are quite complex to manage. My guess is that prices will continue to rise due to supply constraints, and we'll see much more complex optimization behavior.
- The internet is being rebuilt for machines. This TechCrunch article is a little light, but I think the headline and premise are dead on. For a more cutting-edge view on what's happening, read Kin Lan's detailed post. There are multiple categories that now have products explicitly designed for use by LLM agents (Parallel, and Exa.AI for search, for example). This changes a lot, and an excellent preview of some of the things to come can be found in Parallel's founder Parag Agarwal's interview on Stratechery a week ago.
- The Last Technical Interview. Another great Steve Yegge post. This time on hiring software engineers. This has always been hard, but AI is making technical assessment almost impossible. He proposes a mechanism that engages candidates with real work that they are paid for, which then also becomes part of a vetted portfolio. I really like this idea (and, at scale, it would have network effects) - perhaps it's been built, but in the age of vibe coding, hopefully it'll only take a couple of days.
- Salesforce Has an AI Vaporware Problem. Gizmodo covers several reports on the availability of Agent functionality from Salesforce and some other vendors. It's interesting to see this since there are others that report successfully using Salesforce Agents (the SaaStr team, for example). If I were to guess, it's likely that it's very difficult to test agents and be sure they'll perform well when in production, especially in high-safety applications like healthcare. These are complex systems to build and even harder to configure per customer.
- Exclusive: Microsoft is building a super app that combines coding, chat, and other Copilot AI tools. Despite having many AI interfaces, Microsoft may well be deploying another, but it may be the most important. Up until last year, the deployment pattern for AI was to embed it natively into every product. This made sense in a paradigm where users continued to execute all their current tasks, just with "more AI". But Claude Cowork (and to some extent OpenClaw) have upended this paradigm. The new paradigm is a single entry point for all AI use across all tools. What we're seeing now is the fight for the top-level interface to knowledge work, and Microsoft seems to recognize it needs to win this fight and not end up as a software stack underneath Claude CoWork or Codex.
Wishing you a great weekend!
