Felix

building agents, with agents. interpret my own fermata,

[Felix @vcfgdev]

spent the last 2+ years building around llms: infrastructure, rag, multi-agent workflows, deep research products.

across all of that, one thing has become clearer to me: the model matters, but it isn't the whole agent.

the model is the core. the agent takes shape when you build context, action, and workflow around it. that's what people actually interact with.

as more ai products move from copilots toward agents, i've been thinking about where human judgment should stay, and how people can learn to steer them well.

writing some of that down here, and trying to stay grounded while this space moves fast.

[Felix @vcfgdev]

@svpino I stopped reviewing fully AI-generated code line by line earlier this year. Instead, I've been building rules and end-to-end tests around AI agents to ensure that their output stays within predefined constraints before human review and feedback. Burden offloaded.

[Felix @vcfgdev]

@gdb Just wondering: does anyone still miss 4o?

[Felix @vcfgdev]

@rezoundous news: $20 for GPT Plus with 5.6 Luna (max) fast may be powerful enough. some people might not need to pay $200.

[Felix @vcfgdev]

> You need to train the top slice and publish what they build. For everyone else, put the AI in the background of the systems they already work in.
This has been a hard lesson for me as well.

Over the last few months, I worked with a team to roll out AI agents in some companies that were still in the early stages of AI adoption. We used two different approaches. In smaller companies, let everyone try. In larger companies, let department heads to choose a few power users to build and improve skills and then let others use what they build. The latter approach worked better.

The lesson: don't try to make everyone an AI power user since resources are limited. Focus on the people who are willing and able to dig deeper, then let them share what they build.

[vas @vasuman]

https://t.co/haApURmzAN

[Felix @vcfgdev]

@ClaudeDevs i'm reminded of this post https://t.co/N9nbzrz4S2

[Daniel San @dani_avila7]

Claude Code quietly shipped a new Auto Mode slot: hard_deny... flew under the radar but it changes the model

(bookmark this for later)

Been playing with it and it sits next to allow, soft_deny and environment, but behaves differently from all of them

[Felix @vcfgdev]

@thsottiaux How I wish there were no usage limits 🥲

[Felix @vcfgdev]

This is exciting. Combining computer use and browser use could bridge a long-standing gap: when a web page opens an OS-native file picker, an agent that relies solely on a browser automation CLI, like agent-browser I'm using, cannot observe or interact with it, and this often leads to long waits and timeouts.

I had only been experimenting with computer use on Cua Driver before. Just now I ran an experiment: my coding agent with Cua Driver recognized the picker, noticed that it didn't expose an AX tree, and sent an Escape keypress to close it. Now I'm digging deeper into it.

[Cua @trycua]

1/ Today we're bringing browser use to Cua Driver: what we believe is the first extension-free browser use interface built into a unified computer-use driver. Any agent can use exact Chromium tabs and native desktop apps in the same session

[Felix @vcfgdev]

I've been thinking about how to share AI conversations responsibly. A link can preserve the process while your own words should carry the point. https://t.co/M5vm3V6Vtb

[Felix @vcfgdev]

I just recalled the article The Typo Vibe Shift from The Atlantic:

> They want to hear the distinct cadences of a CEO, an influencer, or a celebrity, so they can believe that they are reading something genuine.

Maybe people are just appreciating the human effort and touch behind the words.

[Felix @vcfgdev]

I'd like to share an experiment I carried out about a month ago on turning browser workflows into fast and reliable agent skills, with some encouraging results. It primarily uses @vercel 's agent-browser as browser automation CLI and Node.js as the workflow script runner.

In several internal test cases, it was faster than an agent executing browser actions one step at a time. I also ran an extra test: for the same browser workflow, the skill created with my approach was surprisingly faster than the skill created with the Codex app.

The main idea can be summarized as follows: handle one page at a time, examine the page before acting, and convert actions into scripts.

[Felix @vcfgdev]

@OpenAINewsroom At first I thought it was an iMessage screenshot, but it's not. Wow.

[Felix @vcfgdev]

@dani_avila7 reminds me of the days when I was learning Git and trying to tell the differences between rebase, merge, and checkout.

[Felix @vcfgdev]

I've been enjoying this distraction-free reading setup with @ReadwiseReader in @vivaldibrowser for a while now. Sharing this in case anyone wants a cleaner way to read.

Tab bar and Address bar hidden.
Browser window narrowed and Reader collapses both side panels automatically.
Custom font via Tampermonkey.

Typeface in the screenshot: Fern (V2) Text

[Felix @vcfgdev]

@jdxcode Structure and rules have their own beauty. Some youngsters will always want to understand the gears rather than just care about the output.

[Felix @vcfgdev]

@haydendevs What if the boss actually doesn't know

and that's why they hired you?

[Felix @vcfgdev]

@evisdrenova @pidotdev is already heading in this direction. Minimal harness, multi-provider and open source.

[Felix @vcfgdev]

@lucasmeijer The next milestone is when it ends with "Do you have any questions?" and waits...

[Felix @vcfgdev]

@mitsuhiko Models are becoming more agentic, treating a simple request as a full task, guessing the unmentioned parts, and offering to do more. This might help in coding, but not in writing.

[Felix @vcfgdev]

@zebassembly Agentic scores keep climbing while human judgment remains essential. Someone still has to review and make the decision.

[Felix @vcfgdev]

@dani_avila7 Imported a Claude Code skill into vanilla pi.

skill: "Call AskUserQuestion."
pi: "Call what?"