


Spell casting with AI agents🪄


do u fucking ppl understand where we're at rn
Researchers have used reinforcement learning to teach a 5-tonne John Deere skid-steer loader to perform complex acrobatic manoeuvres autonomously. Trained in simulation and transferred to a real machine, the system can balance on two wheels, navigate rough terrain, spin, and control its bucket before landing safely. The research demonstrates the potential of AI-driven automation for heavy construction and industrial machinery. Video Credit: @leggedrobotics / @ETH Paper: wevlv.co/4kltYys #engineering #technology

most developers would do much better just using one AI chat a time and actually think about what they are doing too many of y'all larp

you cannot automate what you do not understand and to understand, you cannot first automate slow down to go faster i guess

If you have seriously used AI in any real capacity you know just how true this is
The most difficult thing to explain to people is that AI is actually underhyped

For weeks I blamed Antigravity's harness for why Gemini 3.8 Flash (and 3.7 Flash before it) was so slow on coding jobs delegated from Claude Code. I tried rebuilding the harness around it: - Swapped agy's hardcoded PowerShell on Windows for a Git Bash tool. - Replaced agy's two edit tools with one patch tool: agy's edits one block per tool call, so a 3-place edit went from 3 tool calls to 1 - Added batching rules: make independent tool calls together, and run several searches in one bash command (which only raised batched bash searches from 18% to 40%) None of it fixed the slowdown. The real problem was the model. Flash made 1 tool call per turn (~7 s each): across 394 turns, 0 batched 2+ tool calls despite instructions, so every tool call took a full turn. Three more habits made it worse: - Read files in small slices and re-read parts it had already seen, causing 66% of context growth - Ran git diff/status 4–10 times per run to re-check itself - Kept polling background commands with extra tool calls, even though agy reports back when they finish Gemini 4 Hydrogen changed all four out of the box. Across 22 real jobs, about 1/3 of its turns batch 2+ tool calls. Small test on 3 repo tasks (1 run each, medium effort), 3.8 Flash vs Hydrogen: - Time: 979 s → 274 s (~3.6x faster) - Tokens: 491k → 157k (~68% fewer) - Re-reads of already-seen file ranges: 28 → 0 - Tool calls spent polling: 6 → 0 Google's partners saw 2–2.8x faster runs and 12–23% fewer tokens, while for me it was even better.

BREAKING: China says the US should stop fabricating the so-called Chinese spy narrative as a pretext for imposing technology restrictions.

i am sick of the models. you can get the models to agree with almost any architecture proposal, suggestion, research direction, etc. sol, fable 5.2, it doesn't matter. "you're right, your solution is better than mine," its all so sycophantic it makes me sick

Pointed Claude Code at this Gist and one my plugins (grok-delegate) and asked for the same thing for Traycer AI. A few minutes later I was handing work from Claude Code to Gemini Carbon through Traycer.
Sharing 32 best practices for building a plugin that lets Claude Code hand work to another coding agent's CLI. With a delegation plugin, you can: - Send technical writing to Antigravity - Run research on X through Grok Build - Hand code review to Codex - Fan out parallel agents with OpenCode or Command Code for iterative, high-volume work, keeping it off your Claude Code limits Hand these to your coding agent whenever you are building a delegation plugin. Works for Qoder, Cursor CLI, Copilot CLI, MiniMax Code, etc. The gist also includes 3 extra tips for Windows users. Full write-up: gist.github.com/ahmadghoniem… Bookmark🔖 to save for later.

Anthropic has fought open source harder than anyone. They wouldn't be making this move if it weren't for Jensen's current stance. That's no coincidence.
Our initial public offering: ipo.anthropic.com Legal disclaimer: Anthropic has filed a registration statement (including a prospectus) with the SEC for the offering to which this communication relates. Before you invest, you should read the prospectus in that registration statement (available at: sec.gov/Archives/edgar/data/…) and other documents Anthropic has filed with the SEC for more complete information about Anthropic and this offering. You may get these documents for free by visiting EDGAR on the SEC Web site at sec.gov. Alternatively, copies of the preliminary prospectus may be obtained from: Goldman Sachs & Co. LLC, Attention: Prospectus Department, 200 West Street, New York, NY 10282, by telephone at (866) 471-2526, or by email at prospectus-ny@ny.email.gs.com; or Morgan Stanley & Co. LLC, Attention: Prospectus Department, 180 Varick Street, 2nd Floor, New York, NY 10014. The securities may not be sold, nor may offers to buy be accepted, prior to the time the registration statement becomes effective. This post is not an offer to sell, or a solicitation of an offer to buy, the securities in any jurisdiction where the offer or sale is not permitted.

Haiku and Luna got a new contender in browser automation! Meet Meta's Muse Ember. Same price per token as Haiku 5.5 and GPT-6 Luna: $0.10/$0.50 per M. • Model cost per task on @browser_use's Bench V2.1: Ember 5.3¢, Luna 12–16¢, Haiku 23–46¢ • Price jump for long context: Ember at 200K (3×), Haiku at 100K (5×), Luna at 272K (2× input, 1.5× output) I ran it with Browser Use Cloud browsers on an Instabox VM set up by Claude Code with Muse Spark as the judge. 4 of the bench's 200 tasks Ember @ high: • Avg score 74 (100 / 90 / 60 / 46) • Median 9.8 min per task (Haiku high 9.3 min, xhigh 14.2 min) • ~350 tok/s streaming (crazy fast)

Took me a minute to figure out this is about Jensen Huang
A man who pioneered parallel processing cannot be expected to run a single thread

holy moly the dev job market is this cooked? i put out a feeler for candidates got 100+ reach outs in 24hrs decided to talk to a couple of them and lets just say im a bit disappointed. most had no technical ability or engineering mindset or even understanding what its like to approach a problem from first principles. i gave 4 of them a take home project to build a coding harness, and OH MY GOD, one of them used ollama for the backend with some random 8b model from hugging face, the other used AISDK but wrote the whole LLM loop from scratch?!? another one literally told claude to reverse and copy codex, literally runs on codex-server but everything is renamed 💀 4th guy wrote his own agent loop from scratch, but messed up on conversation history and pretty much invalidated the whole cache on every message because he appended the current time to the start of the system prompt and each user message 😂 when i asked why he did this he said "because agents can never tell the time correctly" i said you know if an agent wants to tell the time they can just look at the system time or any time service out there right? he said "yes but this is more convinient" i said "what about the cache?" he said "yes i built a cache where the history is saved" deep interviews are pretty much the only thing standing between you and a horde of non technical vibe coders.

TIL: forking a session keeps the main agent's cache warm since both share the identical prefix, and every hit resets its TTL forks are overlooked compared to subagents tbh and they also get the full context instead of a summary
Subagents are the biggest cause of AI failures In most harnesses: 1. Main agent writes a summary of what it knows 2. Subagent starts from that summary only, and makes assumptions 3. Subagent works and sends back a new summary of what it did Same flaw as compaction: all summaries are lossy. This is how your assistant makes emails you never wanted, or purchases what you told it just to get the price of So I tried something new: checkpoint agents The initial agent has complete context of what needs to be done. And that input is already cached Fork the session where it is. Send the fork ahead to do the task and it will do exactly what the main agent would have done; their context is identical What is even better is when the fork finishes, the cache of the main agent is kept warm by the subagent (they share the same prefix) I measured it on a 137k-token session with Opus 5.5: → Main agent does it on its own: 67 s, $0.38, context +8-17k → Checkpoint: 117 s, $0.68, context +2k → Classic subagent: 123 s, $0.57, context +4-5k But here is what I found: avoiding large input is even more important than keeping cache. And most subagent tasks were done better either in series with retrieval, or split up to be parallel over fresh sessions from the start I do not think there is a place for subagent delegation at all. There should either be a super-agent that orchestrates and does no work, or a long running agent that can retrieve from previous sessions, not a worker that completes tasks partially and delegates the rest to other workers Don't design harnesses based on our assumptions of human behavior

I’m building an accessible app using the #Claude Agent SDK to help people who are blind manage multiple coding sessions at once, without the complexity of terminals or IDEs. It uses their own Claude Code installation and authentication. My dream is to open-source this and help more blind people take their first steps into programming and computer science. @AnthropicAI @ClaudeDevs, could someone help me understand how to make sure this complies with your policies? I’d love to bring this to the community. ❤️ #a11y
