For the past few months I’ve been building AICrier: a signal engine for AI developers. It tracks new tools, model/API launches, agent infra, evals, OSS projects, research, and YouTube content so I can follow the space with more than vibes. I’m also using it as one of my main testbeds for fully agentic coding. I’ll be sharing what’s worth trying, what’s overhyped, and what I’m learning from the build.

Ugghhh, this is tragic. I am proud, however, to have contributed to the BioWare legacy (I, and my team, created the engine behind Star Wars: The Old Republic MMO). You will be missed.🫡

Now this is interesting!

Dumbest thing about @omarchy .... How amazingly easy it is to accidentally get into magnifier/zoom mode. And how ridiculous it is to figure out how to get back out. One key combo is easy to hit by accident the other is almost impossible to figure out even if your life depended on it.

So I messed around with @EntireHQ to see what it was all about. Then kind of forgot about it. Until suddenly, out of nowhere, I find that my PR CIs are being held up by something called "Entire Gate" for some reason. Just spinning indefnintely. I guess they added a new feature specifically to "gate" any of my PRs from ever getting over the finish line. Whatever the intention was, I find this type of behavior obnoxious and Entire-ly unacceptable. If you want to add something new, Entire, then you default it off if it can impact an attached repo and let people opt-in to your new idea if they want it. Ripped that bad-boy out of my repo.

About a week ago, I stood up my dark software factory for the first time after several abandoned attempts at different designs, and evaluating every off-the-shelf solution and rejecting them for one reason or another. I’m gonna have a lot more to say about this, but if I were to some-up my feelings about just how incredibly productive this thing is in a single word it would be: Terrifying.

Grok @bot needs work spaces or folders or some way to separate different teams of bots into their own independent groups. With their own tool identities, and maybe even account to draw their limits from.

Hey @Netlify, our builds suddenly went from about 80s to 210s+ yesterday afternoon. Our own build steps take the same time, but git checkout, cache upload, and some silent gaps before the build starts all got slower. Clearing the cache didn't help. Support ticket's in. What gives? 👀

Some people are thinking about it... some of us are already doing it.

T3Code Stable with Ov2 has landed! Which also means, my PR has finally landed. Why would you care... we... My little addition to this project PR gives T3 Code's agents tools for working in git worktrees, the separate checkouts each agent works in. An agent can hand its worktree off to another session, and it can report what state that worktree is in. That lets orchestrated parallel agents pass work along and keep track of it without stepping on each other. Without this tool, an agent can still run git, but git only knows about files, not about T3 Code. The agent can't tell which session owns a worktree, whether another agent is mid-task there, or how to hand its work to a new session so T3 Code's interface and orchestrator follow along. Agents end up guessing from raw git output, colliding in the same checkout, or leaving worktrees orphaned. With this tool, T3 Code itself tracks who owns each worktree and what state it's in, so handoffs between parallel agents are deliberate and visible instead of improvised with shell commands. You're welcome.

... yup... that was the damnedest thing I've ever seen.

Good idea (for us Convex users)

Welp, OpenAI rolling over and dying. 429s.

sgrep, by @getcontextdev, earns a place in a coding agent’s toolbox. Its strongest result in our tests was reranking searches after narrowing them to a relevant module. The claim that it universally improves coding agents goes beyond what we measured. We tested two real TypeScript codebases: 272 source files and 4,947 source files. We fixed the questions and expected code locations before running the searches. Focused reranking was useful. On the larger codebase, narrowing to a relevant module and piping ripgrep results through sgrep put the target implementation in the first five results for five of six questions. Some substantial improvements: • Course-unlocking logic moved from result 86 to result 2. • A readable-text-color helper moved from result 1,172 to result 1. Those positions reflect fixed, broad ripgrep searches. An agent could also improve them by refining its patterns. Standalone semantic search was inconsistent. On the smaller codebase, sgrep found the target in its first five results for five of six questions. Across the larger codebase, it found none of the six specific implementations we wanted. It often returned related UI, tests or callers. Ripgrep remains the obvious choice when you know the name. Definition-focused searches found all four exact-name controls immediately. Lower token bills are still unproven. Reranking returned 66% less text on our larger-codebase questions, but increased output on the smaller one. We measured retrieval, not complete coding tasks or API spending. There is also a resource cost. Warm sgrep searches took roughly 70–300 milliseconds. Ripgrep was faster. Initial indexing of the larger codebase took 14.7 seconds and peaked around 2.1 GiB RAM. My recommendation: use sgrep for unfamiliar behavior and noisy search results, preferably within a relevant module. Keep exact searches simple. It has practical value. The evidence supports selective use; it doesn’t yet support universal performance claims.

Why is @theo the only one glazing Muse Wonder Full? The rest of the feed is people forehead smacking over it. Even when his own ‘silly’ benches were a flop with it. Which he just dismisses as making games is not coding. Which comes as a shock to my decades of game coding experience.

I don’t know, but taking it all in… I feel @theo coming out so strong on Muse Wonder Full might end up with another GPT-5 style walk-back.