18 months ago @vincent_koc and I were using coding agents to turbo charge the development of automated Agent Optimization, I still think back to these late nights and the fun we were having
We go so much done but it coding agents were so early that we had to pour blood sweat and tears to get things working !
Today I refactor the full codebase in 2 prompts, deleted 2,000 dead lines of code and fixed a good dozen bugs. Things are moving so quickly that it’s easy to get stuck working in a way that is just sub-optimal.
I wasn’t maxing out my Claude Code subscription up until recently but something @theo mentioned changed my perspective. Tokens are not just for building features, they are mostly for testing features. Since then I’ve invested in making every change testable e2e in different environments and both the amount I get done and the quality of the work has skyrocketed !
1
3
51
How did I miss this blog, so many rabbit holes to go down ..
Will keep me busy for a while !
taking a second to appreciate that this knowledge is public
modal.com/blog/trillion-toke…
do you want to compete with @modal on LLM inference, as many before you have tried? here's the manual, glhf
16
A good metric for AI maturity is steering turns per session. The more mature your agentic SDLC, the fewer steering turns you should need
Tracking them is tricky though, we've had good success using Qwen to analyse transcripts
19
Does anyone know where the term "noul" comes from ?
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?
I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev
• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions
AFAICT the shortest path to AI-based economic revolution
25
Not enough people talk about Jev's onboarding flow, it was a joy to get through !
Thought this was a nice touch, makes it very clear what Jev doesn't do upfront
15
At this rate naming the project is going to be harder than building it
We already have ax by @dosco that brings DSPy to Typescript ...
🥳 Excited to start revealing what we've been working on in the last few months. First, we decided to reinvent Kubernetes for agentic workloads with statefulness and fast resumption. Secondly, we are building an agentic orchestrator that will be Google's open agentic orchestrator and runtime. github.com/google/ax
1
87
I’ve been experimenting with remote agents and I haven’t liked any of the options I’ve tried so far so built boxsmith.dev
At its core it includes a persistent machine that is auto-sleeps after an hour of inactivity and always on MCPs
1
1
31
I’ve really enjoyed building on top of @executorsh, it’s really made managing MCPs so much simpler - anyone know of something similar for skills ?
@RhysSullivan any plans for this ?
1
21
“Token reduction is not cost reduction.”
That’s why we invest in behavioural analysis of coding-agent usage.
The biggest savings come from changing how agents are used. More efficient usage patterns can drive >50% cost reduction.
arxiv.org/abs/2607.12161
1
23
UI’s are starting to become underrated, real convo from this week:
> how did you do this so fast, are you using /fast mode ?
> what do you mean ? I dragged and dropped it in the UI, took less than a second
8
I've heard about the drama but anyone have info on the actual solution to Navier-Stokes ? That's what I'm really interested about
1
35
When using loops in Claude Code, don't forget to use an internal of ~55m and enable the 1hr TTL so that you don't break the cache on each loop interation
We are seeing teams save as much as 5% of their entire Claude Code bill with just this simple change !
1
17
This is awesome but also highlights one of the downsides of using CLIs instead of MCPs
MCP servers help with discoverability since tools are loaded in fresh for each session. If you use the gh CLI, Claude doesn’t know this exists so don’t forget to update your rules
Sometimes it's easier to show than tell. We're sure this update will help with that. 👀
GitHub CLI now has a repeatable --attach flag that uploads a local image or video. Reference it inline in an issue, pull request, or comment body.
Available now to all users on GitHub across all plans. 🎉
github.blog/changelog/2026-0…
1
47
Claude Code desktop now imports your cookies from Chrome so you don't need to keep logging in - So nice !
1
42
Does the Claude Code desktop app not support steering ?
Every message I send just interrupts the conversation, maybe there is a flag I'm missing ? @trq212
17








