Mastodawn

RT @DailyDoseOfDS_: Claude Code fully dissected! Researchers from UCL reverse-engineered the leaked Claude source. What they found changes how you should think about agent design. Only 1.6% of the codebase is AI decision logic. The other 98.4% is operational infrastructure. Permission gates, tool routing, context compaction, recovery logic, session persistence. The model reasons. The harness does everything else. This is the opposite of what most agent frameworks do today. LangGraph routes model outputs through explicit state machines. Devin bolts heavy planners onto operational scaffolding. Claude Code gives the model maximum decision latitude inside a rich deterministic harness, and invests all its engineering effort in that harness. The core loop is a simple while-true. Call model, run tools, repeat. But the systems around that loop are where the real design lives: A permission system with 7 modes and an ML classifier. Users approve 93% of prompts anyway, so the architecture compensates with automated layers instead of adding more warnings. A 5-layer context compaction pipeline. Each layer runs only when cheaper ones fail. Budget reduction, snip, microcompact, context collapse, auto-compact. Four extension mechanisms ordered by context cost. Hooks (zero), skills (low), plugins (medium), MCP (high). Each answers a different integration problem. Subagents return only summary text to the parent. Their full transcripts live in sidechain files. Agent teams still cost roughly 7x the tokens of a standard session. Resume does not restore session-scoped permissions. Trust is re-established every ses…

mehr auf Arint.info

#agent #Agent #arXiv #Claude #ClaudeCode #Devin #MCP #medium #nitter #arint_info

https://x.com/DailyDoseOfDS_/status/2065728394084626773#m

Arint - SEO+KI (@[email protected])

RT @DailyDoseOfDS_: Claude Code fully dissected! Researchers from UCL reverse-engineered the leaked Claude source. What they found changes how you should think about agent design. Only 1.6% of the codebase is AI decision logic. The other 98.4% is operational infrastructure. Permission gates, tool routing, context compaction, recovery logic, session persistence. The model reasons. The harness does everything else. This is the opposite of what most agent frameworks do today. LangGraph routes model outputs through explicit state machines. Devin bolts heavy planners onto operational scaffolding. Claude Code gives the model maximum decision latitude inside a rich deterministic harness, and invests all its engineering effort in that harness. The core loop is a simple while-true. Call model, run tools, repeat. But the systems around that loop are where the real design lives: A permission system with 7 modes and an ML classifier. Users approve 93% of prompts anyway, so the architecture compensates with automated layers instead of adding more warnings. A 5-layer context compaction pipeline. Each layer runs only when cheaper ones fail. Budget reduction, snip, microcompact, context collapse, auto-compact. Four extension mechanisms ordered by context cost. Hooks (zero), skills (low), plugins (medium), MCP (high). Each answers a different integration problem. Subagents return only summary text to the parent. Their full transcripts live in sidechain files. Agent teams still cost roughly 7x the tokens of a standard session. Resume does not restore session-scoped permissions. Trust is re-established every ses… <a href="https://arint.info/@Arint/116749319539281425">mehr</a> auf <a href="https://arint.info/">Arint.info</a> #agent #Agent #arXiv #Claude #ClaudeCode #Devin #MCP #medium #nitter #arint_info <a href="https://x.com/DailyDoseOfDS_/status/2065728394084626773#m">https://x.com/DailyDoseOfDS_/status/2065728394084626773#m</a>

Mastodon Glitch Edition

Fedilab Apps Jun 13

I have updated the #Nitter instance used by #Fedilab. You should now be able to follow Twitter accounts again.

Show thread

Wordmark Jun 2

@gilduran.com not the worst way to leave behind #twitter #nitter enables to read it https://nitter.net/PalantirTech?cursor=DAAHCgABHJxOPaM__-oLAAIAAAATMjA0OTg5ODMyMTQxODg5NTY1NAgAAwAAAAIAAA

Arint - SEO+KI May 29

RT @itsolelehmann: this is (genuinely) the best advice i've read on how to build in the AI age. my notes: the question: is there anything left to build, or will OpenAI and Anthropic kill every app? Joe's answer: it depends where you build. he splits it with a Wizard of Oz metaphor. there's the Yellow Brick Road, and then there's the rest of Oz. the Yellow Brick Road is where the big labs already are. anything that gets better as the models get smarter: code, writing, images. build there and you're dead. they own the model, the prices, and the reach. if all you've built is a thin app on top of ChatGPT, they'll just ship it themselves. the rest of Oz is everything else. the messy, specific problems inside one industry, where the model is a small piece and the real work is everything you build around it. think tons of steps, approvals, clunky old software, answers that have to be exact. the tell that this is where the money is: the labs are spending billions setting up their own models for companies one by one. you simply don't do that if the next model just solves it. the obvious worry is that the labs eventually eat this too. but Joe gives a few reasons a focused startup can hold them off: > specific domain knowledge that's written down nowhere. you only get it by doing the job thousands of times. > they use any model on the market, not just one lab's. best tool for each task. > lower costs. cheap models for the easy work, expensive ones only where it counts. > they own the messy legal and privacy rules in fields like healthcare and finance. a general tool can't promise that. it al…

mehr auf Arint.info

#Anthropic #ChatGPT #go #nitter #OpenAI #rest #arint_info

https://x.com/itsolelehmann/status/2060324664795447479#m

Arint - SEO+KI (@[email protected])

RT @itsolelehmann: this is (genuinely) the best advice i've read on how to build in the AI age. my notes: the question: is there anything left to build, or will OpenAI and Anthropic kill every app? Joe's answer: it depends where you build. he splits it with a Wizard of Oz metaphor. there's the Yellow Brick Road, and then there's the rest of Oz. the Yellow Brick Road is where the big labs already are. anything that gets better as the models get smarter: code, writing, images. build there and you're dead. they own the model, the prices, and the reach. if all you've built is a thin app on top of ChatGPT, they'll just ship it themselves. the rest of Oz is everything else. the messy, specific problems inside one industry, where the model is a small piece and the real work is everything you build around it. think tons of steps, approvals, clunky old software, answers that have to be exact. the tell that this is where the money is: the labs are spending billions setting up their own models for companies one by one. you simply don't do that if the next model just solves it. the obvious worry is that the labs eventually eat this too. but Joe gives a few reasons a focused startup can hold them off: > specific domain knowledge that's written down nowhere. you only get it by doing the job thousands of times. > they use any model on the market, not just one lab's. best tool for each task. > lower costs. cheap models for the easy work, expensive ones only where it counts. > they own the messy legal and privacy rules in fields like healthcare and finance. a general tool can't promise that. it al… <a href="https://arint.info/@Arint/116658712233727123">mehr</a> auf <a href="https://arint.info/">Arint.info</a> #Anthropic #ChatGPT #go #nitter #OpenAI #rest #arint_info <a href="https://x.com/itsolelehmann/status/2060324664795447479#m">https://x.com/itsolelehmann/status/2060324664795447479#m</a>

Mastodon Glitch Edition