I think this is healthy for the space, more consolidation. Augment team is top notch

Man I wish I had time to join this debate. But on the point for models being nerfed. What I can say factually is back in Sonnet 3.5 days I had tests I’d run to evaluate laziness. And there would be weeks where it would score just fine and then periods of hours where the model would consistently reduce its response. I had automated tests I’d run every 3 hours 7 days a week. I had a structured output test with increasing items returned in an array. With explicit instructions to always return the fully number requested. It would until it wouldn’t, and then it would add … add more here. I’d be surprised if there weren’t levers of response or reasoning being adjusted possibly dynamically.

Tech drama is the best drama 🍿

Finally had some time to really see what Jev can do, and holy smokes I have an infinite number of use cases to test now.

Opus 5.5 and Astra are both incredible, the downside is I now need to upgrade my Claude code plan too.

Are you flipping my flapjacks? No one saw this coming…

I am shocked

Seeing people build MCPs for agent harnesses they own instead of native tool calls hurts me. MCPs are meant to allow your functionality to live in someone else harnesses.

I’ve been told this is called being a meat proxy.

I had a 3 hour drive yesterday and kept thinking about all these AI doom claims. The one thing I landed on is how in the world would anyone feel good about investing into OpenAI, SpaceX and Anthropic now? It’s either we slow down and valuations need to come way down, or if for some reason there is real liability on these companies that they can hurt humanity the companies would be sued into oblivion. Anthropic would have to completely change its IPO plans now. I also keep thinking about the severe lack of evidence provided other than Jacob’s 10% claim that isnt backed by anything factual. Playing game theory here, if we end up with a big government AI agency responsible for regulation of AI, it could kill open weight models, and make it where government can control what it believes AI can and can’t say. Why would these companies want this?

Tinfoil hat on. What if we are hitting limits and the best way to cover this up is to use safety and slowing down as the driver.

Good thing I know woodworking

Seeing so many non techy people prefer working in Claude Desktop over logging into a website. I think I might of underestimated how quick that would happen. Do SaSS companies just accept this and work on improving their product in ChatGPT and Claude or do they try to compete with their own offerings?

Nothing more frustrating then trying to stop someone from building in a way that you firmly know won't work. You try to redirect the team and help them understand what you've done that proves the approach won't work. But the team charges on, 3 months later the team is just now realizing they should of gone the direction I pushed for. Folks get frustrated because now everything needs to be changed. Its painful watching people running towards a wall and you try to redirect them. I guess some people need to bang their heads a few times.

An engineer called me over the weekend to help him with some problems. I ended up spending hours with him. What I found is he had gotten so reliant on AI that he wasn't even thinking about the customer experience, the product, the outcome. The best engineers will focus on the customer. He was so focused on using AI to automate everything, (he's in qa automation) that he wasn't even thinking about what should be tested or not tested. He was getting beat up on the number of defects that were making their way into customer hands. Its a small startup with less than 10 people there. He showed me his pipeline of work, which basically consisted of taking a PRD from a PM with AI generated AC (which was really bad) and then feeding that AC into Claude to generate a bunch of tests. When the tests pass, he would greenlight the new feature. I gave him a ton of advice, and worked through some ways we could improve the workflow. But the bottom line imo is you do need to have some level of human involvement, even if its validating and just thinking for a few minutes on if the right things are being tested. For example one of the AC was "XYZ modal works" Not to mention i'm super skeptical of QA automation, always have been, and with AI I kinda think we could probably have computer use do what it needs to validate major end to end workflows.

I’m surprised how many people disagree with this. As humans we have limited cognitive load, and as systems expand it’s why we organize around domains. Eventually codebases outgrow what one human can keep in their head. There’s a book called Team Topologies that does a really great job of helping understand how we scale engineering organizations. But you should understand the codebase very highly in your domain. You should understand upstream and downstream impacts. And if your codebase is small, you should definitely understand the entire thing.

codex --yolo Lets see how things finish up overnight.

Every one of my benchmarks models use to struggle on is saturated now. Even some of the fun tests like build a fully playable pool game just works with GPT 6 - Astra. I am so fascinated by what this model can do. By far the best version of this done so far in the 100's of models i've tested. The only thing I could nitpick is the stripes are not positioned well on the balls. Physics are great, it even correctly makes you call a pocket for the 8 ball.

We now have a dragon guarding its eggs, that shoots fire at you. Even a year back AI had trouble making a competent pool game, now we have full playable multiplayer games.

I’ve been building something called Avenrook. Basically I wanted Grok Bot but running locally with any AI model possible. Is it as good as Grok Bot, nope not yet, but its already pretty useful. I wanted a place where I could give a few AI agents a goal and actually watch them work together. Different models, different jobs, their own computers, all in one desktop app. So that’s what I built. Each agent gets its own chat history, memory, model, terminal, Chromium browser, and persistent Linux sandbox. They can write code, run tests, clone repos, research things, browse websites, take screenshots, and create files you can actually use. You can watch their browser and terminal while they work, jump in yourself, or just follow the conversation. The part I’m most excited about is group chats. You put a team together, give them a goal, and the lead breaks it into assignments. Agents can work in parallel, send each other questions, share reports and code, and review one another’s results. If something needs work, it can go back for another pass. There’s a task board showing who’s doing what, what’s blocked, what’s been reviewed, and what’s finished. You can add direction while they’re working, pause the team, or come back later and resume. The app checks that the tasks and final deliverable have passed through review before marking a goal complete. You also get to choose the model for every agent. Run Grok through a cloud provider, put a local model on another agent, and have them work in the same group. It supports OpenRouter and OpenAI compatible endpoints, including local setups like Ollama and LM Studio. Switch a model without losing the conversation. I’ve also put a lot of work into keeping their activity contained. My setup runs the agents in separate containers inside a dedicated Linux VM with no Windows directory shares. Each gets its own files and browser profile. Sharing work between agents happens explicitly through published artifacts. A few other things already working: • Streaming chat with visible tool calls • Web research with citations • Image attachments, screenshots, and image generation • Memory notes you can edit yourself • Individual agent permissions and approval settings • Search across conversations • Conversation exports and downloadable reports • Chat history that survives restarts • Clear history without wiping the agent’s memory or sandbox files • Turn and time limits for unattended team runs It’s running as a Windows desktop app now, and I’ve been testing it with cloud and local models. Still plenty to improve, but watching a team divide up a problem, produce actual files, and check each other’s work is already pretty wild. Next thing I need to do is make it so I can have cloud VM's the agents use instead of local VMs on the local computer. I keep having them get me flagged as a bot, which sucks when you actually want to visit a website they got you flagged on.