Read all 21k characters and drew them out. Slide 3 is the part worth stealing: one interview becomes 6 months of posts.
first auth change i've seen that matters for agent setups.
my agents browse with my logged-in sessions on a mac. they keep working. someone break in and copying those cookies to its own box doesn't
Today we are starting the rollout of Device Bound Session Credentials (DBSC) for vercel.com
DBSC makes it impossible to steal session cookies from a device. So, if you install info-stealer malware (e.g. through a supply chain attack), that malware can not steal your vercel.com session.
Malware can still do bad things while residing on the device. But the abuse we are seeing in practice separates the stealing from the eventual abuse, and DBSC effectively prevents this.
DBSC works by storing a private key in your device's TPM (the same device that the fingerprint sensor or similar biometric devices on your laptop or phone uses) and using that key to sign your cookie. The private key can never be retrieved ever again. Not by us, not by you, not by the malware.
Every website should use DBSC but given how new of a technology it is, we still need to build more experience in rolling it out. In our case there were lots of special cases for things like oauth apps that we needed to resolve. This is also why we are taking a slow approach to the rollout.
DBSC is currently only supported by Chrome on Mac and Windows. We are also working on a kind of polyfill for other browsers. However, the polyfill is less secure and we would love to see broader browser adoption.
There is a related technology called DPoP which is essentially the same but for apps that run on your device. We're in the process of shipping this for Vercel CLI. Ideally, all CLIs and apps would adopt this technology. Luckily rolling out DPoP is much easier than DBSC and it does not depend on browser vendor support.
7
checked my own logs. last 7 days my agents ran 30,642 shell commands and 3,091 MCP calls, and most of those MCP calls were the app's own built in tools
every tool i wrote for them is a CLI. 213 of them. they read --help, pipe the output into grep, done
My friend @stevelacey says AI agents almost never use MCPs if they have it available
I thought the same before which I was against MCP so long
I always thought APIs could do the same already, in the back they're almost exactly the same
Long term I think MCP and API endpoints will just merge into one, both supporting API keys and OAuth account level access
I have been adding MCPs to my site simply for the fact that people expect them now but as Steve says it's debatable if AI agents actually will use them over whatever other way there is to use your site
9
Step 7 is the one I'd steal first: the agent films its own UI, reads every frame and fixes what it finds
the Head of Design at X just open-sourced his taste (a skill that makes your agent's UI look production-grade)
I broke down how it works ↓
Benji wrote down the hundred small decisions he makes when he designs product UI by hand, so an agent can make them too
the skill is built around one sentence:
"it reads as a real product, simplified to what the viewer needs, framed with intent, and every motion has a cause"
it names the 5 ways agent-made UI goes wrong:
> too much, real words in every cell, packed tight
> too little, a wireframe that doesn't read as a product
> boxed, the whole app shrunk into a card like a screenshot
> flat, everything has the same weight
> bland, your brand's style never shows up
then the agent works in 7 steps, and most of them happen before it writes any code:
1. medium: still or animation, where it sits, what size
2. look: light, dark, blueprint or glass, one accent
3. story: 3 to 5 beats
4. examples: start from the closest of dozens of gold examples
5. screen: a few real words, placeholder lines for the rest
6. choreograph: overlapping beats and a loop that ends where it began
7. rehearse: film it, read every frame, fix what it finds
the agent films its own work, a script creates every frame on a sheet and monitors one-frame jumps and text that doesn't fit its box, which a screenshot can't show
it reads your design tokens too, so the UI uses your real colours, fonts and corner radiuses
he built it by giving the same short prompt to 2 new agents with no other context, one with the skill and one without, then comparing the results side by side
when something looked off, he wrote down what was wrong, updated the skill and ran the whole test again, instead of fixing it by hand
works in Grok Build, Claude Code, Codex and Gemini:
npx skills add benjitaylor/skills --skill stageshow
Benji says most results still take him a few rounds of feedback, and the first pass is already a big step up from what an agent makes on its own
YouTube ranks by watch time. HackerTube ranks by how many different months HN people kept recommending a video.
The whole pipeline @karpathy had Claude build, in one picture
1
1
21
Still love Opus 5.5, but Codex Cloud is really impressive. I’m begging you please give it a try.
2
28
tibo really hates people hoarding tokens for the weekend
any day in the next 28 can be a full reset, so spread your usage across the week to take advantage of it
Over the next 28 days, each day we’ll either ship one thing that is a clear improvement and relevant for most codex/work users or ship a full reset. Let the improvements begin.
2
45
I'm honestly split between one long running chief of staff thread and just opening a new thread every time.
memory got good enough that a fresh thread figures out where it is on its own
tested it on my own setup: claude answered 10/10 questions about it from codex's memory file, 0/10 without
I haven’t touched Claude Code or Codex CLI in a while.
The terminal era is over imo. It's the wrong interface for coding agents. Tabs are ephemeral, but context is persistent, and managing 30 tabs is pure cognitive overhead.
I don’t really need an IDE like Cursor either. I rarely navigate the whole codebase anymore.
The new primitive is the agent, not the file.
(Codex desktop app is the best agentic UI for now. But we’re still early.)
3
43
If your subagents just read files and grep all day they don't need to be sol, add this to the codex config
1
21
I has opus babysitting 2 codex agents all night, and when codex hits 0% opus goes to the X asking Tibo for reset
we're so back
2
70
2021 me would sell a kidney for Copilot.
2026 me is annoyed Sol thinks for 13 minutes about a text change.
2
37
unpopular opinion the $500 plan looks cheap by next devday
we all laughed at $200 too and here we are??
2
37


















