Stopped using Claude Code Projects and moved back to a Chief of Staff workflow on my local machine. Claude Projects asks for too many permissions, interrupts often, and switching back and forth between threads is cumbersome. Hope these get fixed once it's out of beta.

Now I have to decide whether to keep Codex by downgrading to the $20 Plus plan or continue with Devin at $20. Devin is good, but I haven't even used the free SWE-2 in days. Its real appeal for me is Cloud, where $20 isn't enough. For CLI/Desktop, Claude already covers my needs. Codex seems more useful as a backup, especially for reviewing Claude's work.

Just got access to Claude Code Projects. Watching a coordinator thread orchestrate parallel threads in real time is a massive workflow upgrade. A few rough edges here and there, but overall a seriously useful addition.

I still have a week left on my Codex subscription, with 3 banked resets, and Devin SWE-2 is free too. But since switching to Claude, I haven’t felt the need to use either. Claude’s usage limits have been more than enough for my work so far.

After a lot of thought, I finally subscribed to Claude Max 20× and decided not to renew my Codex Pro 20× plan. Codex is still a great product, and the recent usage resets and improvements are welcome. But I just couldn't resist moving to Opus 5.5. Let's see how this change works out over the next few weeks.

SWE-2 is now free in Devin for October month. But for Desktop/CLI only. I wish it would have been for Cloud as well.

Yesterday I posted that Sol 6.1 is super slow but today it is working fine. Speed has improved definitely, can't say temporary or permanent?

Unified my Cursor and X accounts on the Plus plan today, only to find Auto-Review disabled in settings. Turns out it relies on third-party models (Gemini 3.5 Flash Lite / Claude 4.5 Haiku) that aren't included in Plus. Without Auto-Review, Cursor is honestly barely usable.

From last 2 hours Sol 6.1 is working on minor bug fixes and not a single PR created. Super slow… Seriously thinking to switch to Claude now…

Now we can check how much storage is taken by Codex on the computer. Mine is 63.8 GB, need to clean up.

Codex Desktop on Linux doesn't yet have built-in computer use, so I decided to experiment with Cua @trycua . Using Cua Driver, I got Codex to detect application windows, capture screenshots, click through settings, and interact with the UI. Getting it working on Ubuntu took some effort. GNOME/Wayland permissions, accessibility access, and window targeting needed attention, but the results have been encouraging. It's still experimental, with a few reliability issues to resolve. But getting Codex to operate native Linux apps without disabling security protections is a good step forward. Hopefully, native computer use will eventually come to Codex Desktop on Linux too. github.com/trycua/cua

It’s been weeks since I last opened VS Code or Zed just to look at code. Coding agents have changed the way I work so much that I rarely feel the need anymore.

After using this I can say it is very useful feature. Much better and accurate predictions than I expected.

Every time things start going well for Anthropic, they seem to score an own goal. Just when Opus 5.5 is gaining traction and people are considering moving from Codex, along comes the Claude for Startups fiasco. First accepting startups without proper verification, then rejecting legitimate ones without giving clear reasons. Such a strange way to lose goodwill when the momentum is finally on your side. @sarahzorah

Devin Cloud is amazing, but over the last 2–3 days, sessions have been crashing frequently. When restarted, they seem to lose track of their progress and jump back several iterations, as if the recent work never happened. Quite frustrating, especially for long-running tasks. @devindevelopers @dabit3

Haiku 5.5 launched yesterday, and our early tests have been encouraging. We've spent weeks evaluating LLMs for accurate document reading, including Gemma 3, Mistral Small 4, Gemini, GPT-6 Luna, and now Haiku 5.5. The challenge is reading exact values without guessing, changing digits, or mixing up information. On a difficult 17-case test, Haiku got 15 correct (88.2%), compared with 13 for Mistral Small 4 (76.5%), 3 for Luna xhigh (17.6%), and 5 for Luna max (29.4%). Haiku high was also around 4× faster and 2.6× cheaper than Luna max in our 30-document comparison. What interested me most was that increasing Luna's reasoning effort didn't improve the results enough to close the gap. These are small, selected tests with different configurations, not a universal benchmark. But for our document-reading use case, Haiku 5.5 looks very promising on accuracy, speed, and cost.

I experimented with Devin Cloud Fusion using my ChatGPT plan. It definitely makes the Devin daily quota last much longer, but my Codex Pro allowance dropped 8% in no time. I then disabled the ChatGPT usage toggle, and Devin’s remaining quota was exhausted within 5 minutes. This makes me suspect that with ChatGPT enabled, Fusion is using Astra heavily as the lead model. Devin should let us choose the lead model in cloud, just like Cloud and CLI. @dabit3 @cognition

Interesting to see how Adobe responds to this…

After experimenting with tens of parallel worktree agents in Codex, I’m back to sequential, one-by-one implementation. For complex issues, especially in live line-of-business applications, linear execution gives me much more confidence and peace of mind. Speed is not everything.