Ramya Chinnadurai 🚀 @code_rams

@DeRonin_ Looks cool, so what does the agents actually do? How do they communicate with each other or how you orchestrates the tasks?

AI has written like a typewriter since day one: one key at a time, no backspace. Google's new DiffusionGemma writes like a draft instead. It dumps a rough block of text, then edits it into shape: 1. Whole paragraphs at once, not word by word. Up to 4x faster. 2. It fixes its own mistakes while writing, instead of being stuck with the first guess. 3. It runs on a single consumer graphics card, not a data center. 4. Free and open under Apache 2.0, so anyone can run it themselves. Why it matters even if you never touch the code: fast, local, and free together means smaller builders can run real AI without an API bill they do not control. The catch: it is experimental. Speed is easy to show off. Whether the writing quality holds up is the part still unproven.

Google @Google

10 hours ago

Meet DiffusionGemma ⚡ Our latest experimental open model (Apache 2.0) that generates text up to 4x faster. Instead of predicting and typing just one word at a time like most language models, it drafts and refines entire blocks of text simultaneously. Here’s how it works 🧵 ↓

95 308 3K 169K 795

@embw_l0x 🤣🤣

Its Fable time!

ClaudeDevs @ClaudeDevs

a day ago

We've reset 5-hour and weekly rate limits for all users. Enjoy Fable 5!

1K 2K 33K 1.9M 2K

Anthropic just shipped the most capable model they have ever released to the public. The clever part is not the raw power, it is how they handled the risk. What Claude Fable 5 actually changes: 1. It is "Mythos-class", a tier that until today only ran inside Anthropic. First time this level is open to everyone. 2. Built to run for days, not minutes. It plans across stages, hands work to sub-agents, and checks its own output. 3. Stripe tested it on a 50 million line codebase and ran a migration in one day that would have taken a team two months by hand. 4. Costs $10 per million input tokens and $50 per million output. Worth it for heavy long projects, overkill for small daily tasks. 5. The safety move: dangerous queries like cybersecurity or biology do not get refused. They quietly get answered by a weaker model, Opus 4.8, instead. Refusing every risky question looks safe but punishes real users. Downgrading only the dangerous edge while keeping full power everywhere else is the smarter call.

Claude @claudeai

a day ago

Introducing Claude Fable 5: a Mythos-class model that we’ve made safe for general use. Its capabilities exceed those of any model we’ve ever made generally available.

5K 14K 101K 48.8M 21K

@pramodk73 Awesome, Congratulations Pramod!

Putting Claude on autopilot is the easy half. The half nobody sells you is the brake. Wrote the honest version: a 7-step stack for running Claude while you sleep without waking up to a mess. The engine, runs it: → /loop, find it in a throwaway session first → promote, session then desktop then cloud, earn each tier → budget, cap tokens and match the model on every fire → /goal, give the loop a real stop condition The brake, keeps it safe: → red-team the plan before the loop runs it → ship-check the output against what you promised → scope permissions by cost to undo 4 steps are the engine. 3 are the brake. You need both. Full breakdown in the article.👇

Ramya Chinnadurai 🚀 @code_rams

3 days ago

x.com/i/article/2063…

2 0 8 2K 7

@degensing For reads - 1 auto retry, then log and alert. For writes - no auto retry. Manual review before rerun. The failure I want to avoid isn't the task not running. It's the task running twice and doing something twice.

x.com/i/article/2063…

Most second brains are write-only. You save notes for years, then never find them again. An agent that can actually work inside your vault changes that: 1. Skills over plugins: Claude learns Obsidian syntax natively, wikilinks, embeds, callouts, properties. 2. It writes .base files for you, real database views across your notes with zero manual setup. 3. The agent can read, write, and reason inside the vault, not just answer questions about it. 4. Files stay plain markdown on your disk, so there is no lock-in and nothing to migrate. 5. 27,000 GitHub stars in days, the demand was already there. A vault you can search is a library. A vault an agent can edit is a teammate.

CyrilXBT @cyrilXBT

6 days ago

THE CREATOR OF OBSIDIAN JUST TURNED YOUR NOTE VAULT INTO AN AI AGENT. Not a plugin. Not an integration. A full agent skills system that teaches Claude Code, Codex, and OpenCode to READ, WRITE, and REASON inside your Obsidian vault like a power user. 27,000 GitHub stars in

51 87 775 58K 2K

This is exactly why I built @TweetsMashApp Bookmarking is easy. Going back is the hard part. Tweetsmash turns your dusty bookmarks into: → AI digests & summaries → Auto-categorized insights → Agent-ready (Claude / OpenClaw / Gemini direct search) with MCP and API support → Export to Notion / Google sheets / notes No more “long bookmark” guilt 🚀 Try it: tweetsmash.app

A single agent running 50 tasks is one person who quits at task 20 and reports it done. A dynamic workflow is a team. One writes the plan. A fleet runs it, one job each. A separate agent tries to break the result. No agent audits its own work. The goal never drifts. Turned a 45-minute job into 8 minutes for me this week. Full breakdown in the post.

Ramya Chinnadurai 🚀 @code_rams

a week ago

x.com/i/article/2062…

0 1 14 12K 39

x.com/i/article/2062…

Problem Claude Code solves with Dynamic Workflows: Default harness = plan + execute in same context window. Works for coding but breaks for long, complex, parallel tasks. 3 specific failure modes it fixes: 1. Agentic laziness - Claude stops at 20/50 items and calls it done 2. Self-preferential bias - Claude prefers its own output when asked to verify itself 3. Goal drift - Long sessions + compaction = original constraints get lost How workflows fix it: - Claude writes a custom JS harness on the fly for the specific task - Spawns separate agents, each with their own context window + focused goal - Can choose which model each sub-agent uses - Can run agents in isolated worktrees - If interrupted → resumes from where it left off Examples from the article: - Mine flaky test root causes (run 50 times, form + test theories) - Extract CLAUDE.md rules from 50 past sessions - Find recurring Slack bugs with no tickets filed - Tear apart business plan from investor/customer/competitor POV - Rank 80 resumes + interview you for a rubric - Rename a model across entire codebase Key line: "Dynamic workflows often use more tokens - think carefully about when and how to use them."

Thariq @trq212

a week ago

Workflows are the biggest upgrade to Claude Code’s capabilities since skills and subagents. I dove deep into it with @sidbid to figure out best practices, examples and more.  I’m particularly excited about the non-technical tasks it enables for Claude Code.

178 364 5K 984K 8K

Pain Point I feel daily is, 1. Switching between browser tabs when running agents across multiple machines wastes time 2. The web version lacks native speed and feel 3. Setting up remote environments (Mac Mini, VPS) makes config syncing difficult 4. Daily Hermes users lose productivity without a polished desktop experience This release directly addresses these pain points. Hermes Desktop benefits: 1. Full native desktop app - chats, skills, and configs come along automatically 2. Remote Gateway mode is fully supported - point it at a backend running on Mac Mini or any server 3. Available on all platforms (macOS, Windows, Linux) 4. Zero migration for existing users - just download and log in This desktop version will be especially useful for client work involving multiple agent setups. The remote gateway mode is particularly valuable since heavy agents run on a Mac Mini in my current workflow. Looking forward to a more stable version soon. Already testing the public preview.

Nous Research @NousResearch

a week ago

The next evolution of Hermes Agent is here! Introducing Hermes Desktop: everything you love about Hermes, now native on your machine. First demoed in Jensen's GTC keynote, it's now in public preview.

1K 1K 13K 5.8M 7K

Chasing a new agent framework every month instead of shipping is the most expensive mistake builders make right now. Karpathy's dead list - what's already fading: 1. AutoGen, CrewAI, general agent frameworks.. demos break on real data 2. Agent marketplaces.. too abstract, no specific buyer 3. Benchmark leaderboards.. optimized for evals, not production 4. Horizontal "build any agent" platforms.. add abstraction, remove nothing 5. Per-seat pricing for agents.. wrong unit entirely The tools that survive: narrow scope, real workflow, one measurable outcome.

0xMorty @0xMortyx

a week ago

Andrej Karpathy: "90% of what AI twitter tells you to learn will be dead in 6 months" 90% of what ai twitter tells you to learn dies in 6 months senior engineers already stopped chasing it the dead list: autogen, crewai, autonomous agent pitches, agent marketplaces, benchmark

65 222 2K 378K 5K

Garry built 540,000 lines of Rails to babysit an LLM. That's the Foxconn factory he's describing. Every retry loop, every validator, every guardrail, a bet that the model will fail. Here's what most builders miss: The old economics: model calls were expensive, code was cheap. So you wrote code to ration the model. That equation flipped. Models are cheap. Models write usable code. The architecture should flip too. The new artifact: 1. A markdown file with the intent and the skill 2. Minimal code for the parts that must not hallucinate 3. A test That's it.

Garry Tan @garrytan

a week ago

x.com/i/article/2061…

82 80 966 523K 2K

This week I hit rate limits faster than expected. Turns out it was a bug, not my usage. Anthropic just reset limits for all Pro and Max users. Good reminder: when something feels off with your tools, check if it's actually your fault. Sometimes it isn't. Its time to burn more token! 🚀

ClaudeDevs @ClaudeDevs

a week ago

We've reset 5-hour and weekly rate limits for all users on Pro and Max plans. We fixed an issue that caused some Claude Code sessions to spawn excessive parallel subagents, burning through usage faster than expected.

1K 1K 21K 2.5M 1K

My setup uses cloud Claude (Anthropic API) for the intelligence layer. But for local models, I've been testing with a client setup on Ollama. For 16-18GB RAM, what's actually working: Gemma 4 - Google's latest, punches above its weight at this RAM range. Good for summarization and note-linking. Kimi 2.6 - Moonshot AI's model, surprisingly strong reasoning for its size. Works well for daily briefings and pattern-finding. Both run via Ollama subscription - easiest way to get started. Just ollama pull gemma4 or ollama pull kimi and point your setup at localhost.

x.com/i/article/2061…

Most second brains are graveyards. You save 500 links and reopen none of them. This setup works because the human does almost nothing: 1. Capture runs on autopilot. Articles, podcasts, voice notes land in one vault without you filing them. 2. One vault, every input. Nothing scattered across 6 apps. 3. A job runs at 6am, before you open your laptop. 4. The value is the connections between new notes and old ones, not the storage. 5. Output is 3 decisions, not 300 notes. It narrows to what you act on. The real lesson is the architecture: a second brain only pays off when something reads it for you every day. Saving is easy. Synthesis is the product.

Ramya Chinnadurai 🚀 @code_rams

a week ago

x.com/i/article/2061…

1 1 6 1K 17