Who to Follow on GitHub?
Full Teams
(without Anthropic, OpenAI, Google Gemini, and other top tier systems - you could find plenty of agent skills and tooling there)
TypeSafe AI - Jev tools, skills, and system-level agent tooling
AnomalyCo - OpenCode and the broader coding-agent ecosystem
Browser Use - browser agents, computer use, and web automation
People Inside Strong Teams:
bcherny - Claude Code tooling and developer workflows
NTaylorMullen - skills, skill-creator, and agent capabilities for Gemini CLI
thdxr - OpenCode, agent UX, and coding workflows
cpacker - persistent agents, memory, and long-running state
kieranklaassen - reusable agent workflows and engineering skills
tmchow - coding agent skills and workflow tooling
Independent Authors
Simon Willison - LLM utilities and practical developer tools
David Zhang - Jev, agent skills, and compact AI tools
Cole Medin - context engineering, memory, and agent packs
Joey Kudish - Jev browser, MCP, and agent tooling
Independent Teams
ContinueDev - coding agents, checks, and reusable workflows
Mem0 AI - agent memory, MCP, and OpenMemory
Roo Code - agent UX, plugins, evals, and remote control
Claude Code engineer explains Anthropic's new approach to setting up and developing AI agents in an interview
Thariq Shihipar explained that models are “grown, not designed” - their behavior forms during training and changes as they develop
before:
a new project started with CLAUDE.md, where they wrote rules, structure, commands and limits for the model
now:
“Right now it might be better to start a new project without a Claude.md”
“If you see very repeated failure modes, you add them to your Claude.md”
so a new project starts without a prepared file with instructions for Claude
repeated failure modes become permanent instructions
39 FREE styles for video generation that you can fully control
what if your coding agent could build an video animation which you control by code?
this video took me 2 minutes to prompt, btw
save this not to lose it
repski + list of divided styles below
Make Claude Code read most of the code as images and save 60% on usage without losing quality
works with tasks like:
-> Fix an issue that requires reading many files
-> Understand or analyze a large unfamiliar codebase/repo
-> Run tests and investigate failures
-> Debugging with large logs / test output
-> Long coding sessions with a lot context
the tool turns large chunks of text into images
when Claude reads the images through vision, the context uses fewer tokens and costs less
code that changes remains regular text but in image format
installation below
repo - github.com/teamchong/pxpipe
practical experiment -
1. Check Node.js
// node -v
Need Node.js 20.19+.
if you don't have it: nodejs.org
2. Claude Code
// npx pxpipe-proxy warp -- claude
Then use Claude Code normally
3. Codex
// PXPIPE_MODELS=gpt-5.6-sol npx pxpipe-proxy warp -- codex
Sol is opt in, so the PXPIPE_MODELS part is required for compression
4. Cursor Agent
// npx pxpipe-proxy warp -- cursor-agent
5. Check that it works
Open:
http://127.0.0.1:47821/
The dashboard shows:
- what was converted to images
- tokens saved
- active model
If the dashboard shows conversions and savings, it's working
HOW TO MAKE YOUR AGENT GO THROUGH 10,000 JOB POSTINGS OVERNIGHT
Pinloop is a huge, constantly updated database of job postings
With a single command, you can give your coding agent access to it
letting the agent go through thousands of jobs from dozens of sources based on your criteria
What bout pricing?
Free plan: search and filters, up to 5 new job postings/day, 150 AI judgments/month, and 50 semantic searches/month
Pro plan: 1,500 job postings/month + 1,500 AI judgments/month + unlimited semantic search.
You get autonomous routines/watches that keep searching and evaluating jobs even when your computer is turned off
In my opinion, the best way to use it is as a filter that goes through new job postings every night and gives you a ready made shortlist in the morning
If youre not sure whether you need it, I’d recommend starting with the free plan
coming to pro if it actually works for your workflow
repo - github.com/pinloop-ai/pinloo…
page - pinloop.ai/
CLI install:
1. npm install -g pinloop
2. pinloop welcome
How Claude quietly eats 10% of your usage through prompt suggestions
this feature predicts your next command and shows it as gray text in the input field
claude makes a separate model call that use the context from your current session
So the longer your session - the more usage these suggestions can consume
-> To turn it off, type /config, find "prompt suggestions" and switch it to off
3 ways companies use Jev in their pipelines - and what kind of boost it really gives
Jev came out on September 15. In the first 24 hours, almost 13% of paid AI Gateway teams already used it
Vercel called it the fastest adoption of a new model launch in Gateway history
Here is how companies have added Jev to their workflows 👇
Browser Use is infrastructure for autonomous browser work through AI agent pipelines.
Jev's role - it chooses the next action and the element to use, based on the probability of each option
The runtime creates the allowed set of elements and actions in advance, and Jev chooses from them
After that, a verifier checks the result separately
The test used the same Google Flights task.
They ran the normal version 3 times and the Jev version 3 times, with the same models and settings
After every run, they checked if the task was completed correctly.
The results are below
What boost did company get?
➤ −25% median runtime: 9.45 s - 7.09 s
Jev chose the next action faster instead of using a heavier decision process
➤ −90.8% browser protocol calls: 1,092 - 101
The agent called the browser protocol much less and did less repeated reading and checking of the page state
➤ 3/3 successful runs in both versions
The speed boost did not hurt the result. Both the baseline and the Jev version completed all 3 runs successfully
Vercel is a cloud platform for building, deploying, and running web applications and AI systems
fx is their small coding agent that works with code and runs terminal commands on its own
In auto mode, fx can take actions by itself, but sensitive or unclear actions go through a safety reviewer
Jev's role is to act as a fast safety classifier before the agent takes a real action
Jev is mainly used for semantically unclear actions, where a simple allow/deny list is not enough
btw Vercel moved beyond web development a long time ago
What boost did the company get?
5–18× faster safety review - Jev checks commands much faster before the agent runs them
Up to 18× faster p95 latency - the biggest speedup shows up in the slowest checks
Higher accuracy than GPT 5.6 Luna on Vercel's internal safety benchmark
A GitHub repo stack for setting up and working with Jev
1. TypeSafe - teaches your agent how Jev works, when to use it, which command to choose, and when Jev should be called
a must-have skill tbh
2. SkillRanker - automatically picks the right skill for your agent
If ur coding agent has a lot of skills, Jev checks which ones fit the current task, gives each one a probability, and picks the most relevant one
3. Supercov - helps a coding agent understand where it should spend its next step
it finds weak spots in your tests and code and tells the agent what it should improve next
4. Jev Browser - reduces the amount of context and expensive LLM calls used for browser tasks
It connects to your coding agent and lets Jev handle cheap browser actions instead of using the main model for every click or small decision
5. Jev Guard - checks whether an action is risky before the agent runs it
evaluates whether the action is dangerous and whether the user actually asked for it
interview All IN: Jensen Huang
key quotes:
1. Criticism of AI catastrophe predictions
"What does that mean, 10% of extinction? Nobody knows how to explain that to the average
person?"
--> “We shouldn't, because it's made up”
--> “We ought to just keep track of all that… and remind us that those predictions are inconsistent with ultimately America winning the AI race”
2. What AI regulation should look like
--> “Regulation should solve actual problems. Root cause the problem from an engineering perspective, implement and institutionalize technology, methods or processes to make sure that we don't let it happen again, and use independent auditors or evaluators"
3. How AI companies should be managed
--> “These companies really ought to be built the way that we used to build companies, which is in silence.”
4. China as the largest source of open source
“Probably the vast majority of the world's contribution to open source today is coming from China.”
“When you download one of these Chinese models... it's now yours. Whatever you want to do with it”
5. What the “AI race” really means
--> “The race is really about who exploits the technology best”
6. AI as a new industrial revolution
“This is a new industrial revolution”
“We power anything, we can find anything. Now with AI, we can ask and know anything.”
“This industry isn't just about the model. It's not just about the chips.”
“It's mostly about the applications on top.”
“It's mostly about the infrastructure layer, the data centers... construction, electricity, power generation.”
7. NVIDIA’s strategy: move up the stack only as far as necessary
“Our strategy is: go up as far as we need to and as low as possible”
“We go and we invent all the technology necessary as far as we need to, and then we let a thousand flowers bloom.”
8. China and advanced chip manufacturing
“The way to think about China is: really good at high-volume production.”
“This is just a matter of time.”
“For Nvidia I've got to think about what happens next decade and decade after that.”
“Two or three years is just a click. It's nothing.”
9. Slowing down AI is the wrong strategy
“We got to speedrun.”
“Slowing down is definitely the wrong strategy.”
10. Superintelligence already exists - in narrow domains
“I think we're there, too.”
“When you take a narrow segment... my self-driving car, I don't want you to make me an omelette. I just want you to drive the car.”
“It's better than a human.”
“Synthesizing proteins, doing virtual screening of proteins - we're already there.”
Day 1 of building useless browser features:
Dragon Johny
-> He usually chill on the post button
-> Click on him: he’ll follow your cursor for a while, then return to his spot
-> Click on a post: he sits down and reads a book
-> When you start writing, he moves to another spot and falls asleep
rate how cooked I am from 1 to 10 pls