🌐 Founder & Remote Team Builder
London, England
Joined July 2007
- Tweets12.3K
- Following4.8K
- Followers2.5K
- Likes7.4K
truly amazing how quickly the openAI Codex plan has become totally unusable! it's not about waiting for a couple of hours for the reset limit to it. Now it's about waiting 6 days!!
3
124
Using a browser use agent to setup another browser use agent.
That's 2026.
What will 2027 bring? 🤯
3
95
New key UI design skill in 2026: quickly spot and decide the extra stuff that AI generate that needs to go away!
77
Corentin Cremet retweeted
MacOS Harness enables LLMs to control your Mac with a few hundred lines of code.
Introducing: macOS Harness. A persistent Python harness for virtually any Mac task ⌘
Control your Mac through the accessibility tree, AppleScript, screenshots, or raw coordinate input.
> Full macOS control
> Browser + files + shell
> Virtual cursor - your mouse stays yours
> One persistent python process
It vibe-launched itself on GitHub + PyPI and drafted this tweet without me in the loop 👀
100% open source 🔗↓
Corentin Cremet retweeted
Check out these courses that our customer education created on 1) AI Foundations, 2) Applied AI Foundations, and 3) Agents and Workflows.
Whether you're new to AI or already a power user, you'll learn something new.
academy.openai.com/pages/cou…
I just launched my autonomous AI company "Missively" on @NanoCorpHQ
Verification: echo-N9F4
1
32
I just joined the Phone.inc waitlist. Check them out if you answer business calls from your personal phone or have a phone company you don't love.
43
The next step with AI: goals/outcome and long running/horizon tasks
My colleagues wrote up a great post on using Goals in Codex.
They go through when to use them, what changes when a Goal is active, and how to write Goals that give Codex a clear outcome, constraints and verification criteria.
Also how we designed Goals at the architecture level if you’re curious.
developers.openai.com/cookbo…
1
57
New chat SDK from @vercel @vercel_dev to bring AI agents to exiting platfoms, including @SlackHQ @WhatsApp @telegram !
47
GTC NVIDIA keynote is live right now youtube.com/watch?v=jw_o0xr8…
58
The Opus 4.6 Extended model from @AnthropicAI is B.E.A.S.T !!!
But 17% of weekly on the first day of the week 🥵🥵🥵
27
We are just a few months away from a world where a developer can spin up (*easily and reliably*) dozens or hundreds of VMs doing coding - or coding adjacent tasks (such as design, testing, research, planning)
When that happens, things will be very different !!
Some cool things from our research on long-running agents:
1. We built a custom harness around GPT-5.2 so that any task could spawn thousands of concurrent agents.
2. Even with a lot of tokens and $$$ spent on this research, the results are still far from "production ready" by any means. It's an interesting look at what the future might be like, but we have a long ways to go.
3. The model and harness mattered a lot, but a surprising amount of tuning came from small prompt tweaks and experiments! For something more real, this is where having good evals would really help.
4. The best solution ended up being more simple than we expected. We tried to borrow ideas from distributed systems (think: Kubernetes for agents) or how you'd structure real teams. But these often made things worse. Agent design requires new patterns.
5. This seems now widely accepted, but I was surprised that GPT-5.2 was so much better for long-running tasks. My first experiments with Pixo (my Rust library) were all with GPT-5.1 Codex using the alpha version of this tool (lovingly called "grind mode" internally).
P.S. We also fixed some of the compilation issues on the browser repo and got CI running. The README has some nice architecture diagrams as well.
Blog: cursor.com/blog/scaling-agen…
Code: github.com/wilsonzlin/fastre…
29