Weekly Links: Humans for Tokens, Headless Software, and GPT 5.5

This week, Microsoft gets into Agent Hosting, Codex comes to ChatGPT in the form of shared Agents, and AI reaches elite level in table tennis.

On to the main stories:

• Microsoft and Meta announce large staff reductions as they spend big on AI. This seems to be "humans for tokens" in the most elementary way, but I think most of the media narrative has it backward. Microsoft, Google, Meta, and Amazon see the coming crunch in compute infrastructure, and how much leverage they will have depends on how much data center capacity they have. So human, for now, are more expendable. Part of the story may be that "we don't need them," but I suspect more of the story is "we need the cash to give us as big a position as possible in the crunch".
• An Interview with Google Cloud CEO Thomas Kurian About the Agentic Moment. This interview is a great lens through which to view Google's Next Keynotes this week (available here). I can sum it up in three phrases: data governance, agent harnesses, and owning the whole stack. Google has been out of the AI news for a while, with Anthropic and OpenAI's Codex making the "Agent" news, but they have a strong hand. The missing piece is data governance, and Google is betting that this is compelling for organizations that are scared of letting AI run rampant over their digital content. Google is also clearly bringing to bear its cloud infrastructure knowledge and trying to extend its reach to other clouds. Ben Thompson repeatedly tries to dig into whether Google has gotten good enough at the agent harness problem, but does not get much response. Indeed, Google's answer to Claude Code and Claude CoWork is probably the biggest question that needs an answer.
• Robot runner handily beats humans in half-marathon, setting new record. I think these robot capability stories often get lost in the noise. Seeing a bipedal robot able to easily outrun the human half-marathon record is a huge feat, though. It's the gateway to incredible feats in a few years. It also stings a little if you are one of millions of humans working on their half-marathon PB every year.
• Software will be Headless. I retitled this Aaron Levie interview on Harry Stebiings 20VC podcast (Spotify version is here) because I think this is the true breakthrough statement in the interview. It's a great interview and worth listening to in order to understand the software business going forward. Salesforce's announcement this week of their headless 360 mode got almost no coverage, but I think it's extremely significant in terms of where SaaS is headed. Unfortunately for most SaaS vendors, it seems likely that headless mode might ultimately be just a step along the path to obsolescence. For that not to be the case, they'll need an extremely strong process flow that survives being out of sight and mind of the user.
• OpenAI releases GPT-5.5, bringing company one step closer to an AI ‘super app’. OpenAI seems to time their releases to crash Google's keynotes, but aside from that, it's interesting to see the emphasis on "fewer tokens for the same result". My expectation is that token cost will rise substantially in the next 6-18 months as compute capacity constraints bite. OpenAI and Anthropic, in particular, have generous subscriptions without which some usage would be prohibitive. They are already rolling these back in places. Smarter models using fewer tokens might help, but use cases like agents running overnight on coding tasks will continue to push token usage up radically. If you've working on a vobe coding app, you might want to finish it quickly before prices go up!

Wishing you a great weekend!