Software developer since 1998. Member of #VegasTech and engineering manager / sr. architect at @codingscape (tweets are my own)
Joined November 2007
- Tweets7.9K
- Following929
- Followers651
- Likes4.2K
I wholeheartedly agree. My factory falls apart when it's not using Opus. Has anyone had great success using a factory with something other than Opus in roles like PO or exec?
Astra is a goddamn idiot compared to Opus.
I’ve ran out of opus tokens so I switched (in situ) my product owners to Astra. I then asked it to change a couple of config files and the ensuing conversation is another level of infuriating retardedness.
Same guidance and skills and exact same position in agent topology. Opus would have done it in one prompt and not needlessly overcomplicated it.
Can’t wait until Claude resets.
The Coding Agent Index now grades its coding tasks on whether a maintainer would merge the change. Fail one blocking check and the attempt scores zero.
20 of 21 model + harness setups scored lower. That's the bar a real review gate has to hold.
artificialanalysis.ai/agents…
New in my Granola replacement: the desktop app now syncs my meetings to my own server automatically, and it runs on Linux too. On Linux I've only tested processing so far, not mic capture. georgediab.com/posts/2026/us…
Anthropic went from about 2,500 merged changes a day in May to over 30,000 in September, then replaced git on its monorepo. Nearly all its clients are agents and CI jobs, which are always online. They say it isn't a git replacement for anyone else. anthropic.com/engineering/re…
George Diab ❗️ retweeted
hey. here's a little demo of the most meaningful improvement to my productivity since speech-to-text.
I wanted Granola without the subscription, so I had Codex reverse engineer how it works and build my own. @reatools gave the agent a method: ask specific questions and write down the answers. First commit came about 91 minutes in.
georgediab.com/posts/2026/us…
My agent instructions say "set -o pipefail on all verification chains." I'd forgotten why. Engram traced it to an August session where my agent admitted its `&&` chain hid a failing test, so a red commit landed. The lesson became the rule.
github.com/clickety-clacks/e…
My Linux laptop had two months of agent sessions that nothing indexed. I backfilled 192 of them into Engram in 4 seconds. Now any line of code on that machine can be traced back to the session that wrote it.
github.com/clickety-clacks/e…
I told people they didn't need a dark factory. I'd now tell them to aim for one, and run it like a Michelin kitchen instead.
Agents cook. You don't taste every plate, you sample. When the same mistake shows up twice, you fix the kitchen with a skill or a lint rule, not a PR comment.
Credit to @poteto for the kitchen, and if I'm honest... @bffmike, who came up with the analogy first!
georgediab.com/posts/2026/th…
I had it reverse engineer Granola. Wow that worked super well. REA reverse engineered Granola and I asked it for a clone with a couple tweaks I wanted and got what I wanted after an hour.
X-ray vision for agents.
Binary reverse engineering.
I cannot wait to try this github.com/morluto/rea
X-ray vision for agents.
Binary reverse engineering.
I cannot wait to try this github.com/morluto/rea
Excited to play @SinBinVegas Binball again this year!
Team name: Mitch Better Have My Money
zeffy.com/en-US/ticketing/sa…
I watched the DevDay keynote this morning. Three things are still in my head.
Ultrafast is impressive, and it is priced like it knows that. It is not a new model. It is a speed tier on GPT-6 Astra: up to 8× in Codex (about 300 tokens a second) and up to 6× in the API, in ChatGPT Work and Codex on Pro 500 and Enterprise. Short-context Ultrafast is $60 in and $300 out per million tokens. Standard Astra is $10 and $50. Same model, six times the money, for speed.
What I liked more than the speed demo is ChatGPT as a login and a plan that follows you. Sign in with ChatGPT, and Plus and Pro usage can count across 16 partners, including Devin, Notion, and Vercel. @ChatGPT also shows up in Slack and Teams. Not another window. Your account and your allowance, inside the tool you are already in.
Dots is the exciting one, and the one where I already have a head start. Always-on, its own cloud computer and browser, plugins into thousands of apps, Slack and Teams, and you set what needs approval. The first one is included on Pro and Business Premium. I have been living a version of that desk with Grok Bot, and for the work I actually do, Grok Bot is ahead of this right now.
If you are budgeting agents rather than demos: GPT-6.1 Sol is near-Astra coding at about a fifth of Astra’s standard token price.
openai.com/index/devday-2026…
Just rewrote my global CLAUDE.md after reading Anthropic's Opus 5.5 playbook (by @addyosmani).
Before: "Enter plan mode for ANY non-trivial task" and "Check in before starting implementation."
After, lifted from the playbook: "Stop and ask only when you can't continue without me, or before anything destructive." Plan mode is now for real forks, like architecture calls, not step count.
I'll report back after a few long runs.
Anyone else already flipped theirs? What broke?
claude.dev/blog/getting-the-…
Checked 19 quotes my agent pulled into two notes, character for character, against the source pages. 19 exact. One page was openai.com, which 403s curl and times out WebFetch. Exa's contents endpoint (exa.ai) got it in under 3s for $0.002. I'll take it.
Tried Muse on Mac this morning (ai.meta.com/muse/). Feels Grok Bot–shaped at first (x.ai/bot): Mac app, opt-in access, ask before sensitive stuff.
Asked it to create an agent to watch iMessages from my wife. It said it could — then made a watcher under the same Muse. One assistant. Not a teammate with its own chat.
Desktop cleanup went well: 36 → 10, smart folders, left my keepers alone. Permission prompts (branded “Beacon”) showed real paths. “Allow all once” is per batch, not per job — so you click through waves of file moves.
Asked it to trash some configs, denied the prompt. It stopped. That’s the trust test.
Same silhouette. Different org chart.