Decebal | Rust + Move Engineer ⚙️ @ddonprogramming

@steijnpelle haha she's always got the inside scoop before everyone else. gotta keep notifications on for katie at this point! 📢🔥

@steijnpelle quality over everything. erik knows exactly what he’s doing with that lineup. definitely going to be the new benchmark for how firms handle content. 🤝🔥

@jgreze The Granola + TownAI combo is actual magic for follow-ups. Nothing beats having an automated list of every random promise I made during the day. It’s a total lifesaver.

@jgreze @louismorgner always appreciate when leaders actually keep it real instead of just recycling pr scripts. the bit on sf was spot on. definitely worth the watch for anyone in the space 🫡

@tom_doerr the interview resources in here are actually solid. hard to keep up with how fast things move so having a curated list like this is a lifesaver. thanks for sharing! 🛠️💡

@WillManidis it’s the classic "institutional inertia" trap. they'll grill you on an AI model's 2% error rate while ignoring the absolute chaos of paper records and bad handwriting. the gatekeeping is definitely crumbling though. 💊📉

@FactoryAI this is actually huge for jumping into a new repo without that first hour of "where does anything live" confusion. the ui looks super clean too. nice work on hitting 100 already! 🚀

Launching Rust & AI Weekly: a vetted sweep of where Rust meets AI. Issue #1 — 22 crates mapped by verdict (Adopt/Trial/Assess/Hold): ◆ goose — Block's Rust agent, ~35k★ ◆ qdrant + lancedb — vector DBs Netflix & HubSpot trust ◆ rmcp — the official Rust MCP SDK Every entry has maintenance + adoption signals, so you're not bookmarking something that died last summer. Read it + explore the live radar 👇 decebaldobrica.com/blog/2026-06-1…

@nico_laqua this is spot on. people spend so much time prompt engineering a "perfect" v1 that they never actually launch. shipping ugly code and fixing it live is the only way to actually learn something.

@tom_doerr this is huge for dev workflow. getting that <5 min feedback loop on solidity changes directly in the ide changes the game for catching silly mistakes early. pashov group knows their stuff 🫡

@tom_doerr this is super clean. having pashov's logic baked directly into claude code is going to save so much back-and-forth during dev. definitely checking out that x-ray skill tonight 🔍🔥

@DataChaz @karpathy Markdown and YAML frontmatter is the way to go. Simple is better. It's awesome to see this formalized because proprietary silos are a nightmare for agents. Finally, a standard that won't break when the next model drops. 🚀

@AlfieJCarter AGENTS

The "9-step loop turns Claude Code into a senior engineer" take is half right. A loop is control flow. It assumes free context, perfect memory, and that an in-session reviewer is independent. None hold on a real codebase. The substrate is what compounds: decebaldobrica.com/blog/2026-06-1…

This is the best thread on Claude Code I've read in a while, and I still think it's only half the story. The argument: the model was always capable; the difference is the loop — explore, plan, build small, enforce with hooks, test, review with a second agent, fix, ship. All correct. Most people don't do any of it. But a loop is control flow. It tells you what order the steps run in. It says nothing about whether each step can actually execute on a 400-file repo, across a week, without the agent running out of context or forgetting why it made a decision three sessions ago. The 9-step loop quietly assumes three things that are false on real work: 1/ Context is free. Explore + plan + build + test + review + re-review is 6+ full passes over your code, every task. On a real repo that's a token bonfire — and it hits the context window before it hits "done." 2/ Memory persists. Plan mode and CLAUDE.md live for exactly one session. The moment you /compact, the why behind every decision is gone. The agent re-explores the same code tomorrow. 3/ An in-session reviewer is independent. A review subagent runs on the same model, in the same session, on your token budget, against the same blind spots. It's the same engineer in a fake moustache. The loop is the floor. Here's the layer that makes it actually run: → Token economy: caveman (trims the agent's own output ~65-75%) + rtk/Rust Token Killer (filters terminal output before it hits context). The loop is only affordable if each pass is cheap. Most people abandon the discipline because it got too expensive — not because it failed. → Durable memory: chronis, an event-sourced task CLI — the plan becomes a replayable task graph that outlives the session, with TOON output that's ~50% fewer tokens than an ASCII table. Plus AllSource Prime, a knowledge-graph memory layer that remembers what the system is and why, across sessions. (Both are mine — disclosing that up front.) Plan mode is short-term memory. This is long-term. Senior engineers are mostly long-term memory. → Independent review: CodeRabbit, on the PR, outside your session — different system, doesn't share your agent's blind spots, doesn't spend your context budget. That's what "second reviewer" actually means. → The step the loop doesn't have: pixel-perfect verification. Tests assert behavior, LLM review reads code — neither looks at the rendered pixels. A pixel-perfect overlay shows the design vs. the running app to the pixel. For UI, that's where "done" is earned. The honest part: for a quick fix, the bare loop is the right amount of process. The stack has real costs. The thread's primitives — hooks, plan mode, subagents, CLAUDE.md — are exactly where all of this plugs in. I'm not replacing the loop. I'm giving it a substrate. The model was never the bottleneck, and neither is the loop. The substrate is — tokens, memory, independent verification. Build the loop. Then give it a stack. Full write-up, with the loop wired into the stack step by step: decebaldobrica.com/blog/2026-06-1…

@coreyhainesco Love that you're already jumping on OKF. Everyone's obsessed with llms.txt right now, but getting the metadata structure right for agents is the real long game. Super helpful update.

@deedydas absolute insanity. the 0.4 year median tenure is the wild part—the "last mile" hires basically got a nice bonus while early folks are buying islands. insane run for the cursor team.

@AnatoliKopadze managing 10k agents solo sounds like a fever dream but that's where we're headed. the "junior dev" role is basically becoming an orchestration lead overnight. wild shift to witness.

@Rasmic The amount of time I've wasted trying to hack together what Fable does out of the box is embarrassing lol. It’s one of those tools that makes you wonder why we did it the hard way for so long.

@DataChaz this is the dream lol. actually being able to close the laptop without the process killing over is huge. the "wake up to a PR" workflow is finally starting to feel real instead of just hype 🤝