Sascha Kuhlmann

I'm going LIVE in 90 minutes, you can watch here youtube.com/live/-0p7kEeVK… or here on X x.com/bernieblume

@jeanpierresilva @bryan_johnson Yeah 27mg of Toad is a HUG dose …. 27mg of pure 5-MEO-DMT is nirvana…. 12-15mg is the standard breakthrough dose

@bryan_johnson Maaaaannnn you didn’t use the real Toad!!! Not the same thing man! Am I wrong? That looked like the synthetic kind!

@techxutkarsh @david_ortecho @ViralityEngine Send it to me please

@david_ortecho @ViralityEngine Sure David. Please follow me so I can dm you

A senior Google engineer just dropped a 421-page doc called Agentic Design Patterns. Every chapter is code-backed and covers the frontier of AI systems: → Prompt chaining, routing, memory → MCP & multi-agent coordination → Guardrails, reasoning, planning This isn’t a blog post. It’s a curriculum. And it’s free.

@djrhails @vincent_koc @openclaw how do you override tools ? I created a plugin to block certain tool calls ...but overriding them is not currently github.com/coolmanns/open…

A few ones: - Ability to access the callGatewayTool - Make it possible to override existing tools - Ability to inject messages (e.g. Ralph /loops are super painful at a plugin level) - Better guidance around getting types to work when developing locally - think I’ve got a fix, but feels clumsy. - Agent end hook - needs to be before the agent fully exits the loop so it’s possible to make it continue.

Looking for input / feedback for anyone developing plugins or looking to develop plugins on @openclaw - What hooks/features are you looking for? - What is stopping you from plugins? - Any key pain points? - How do you test when developing? DM's open or drop comments here :)

@burnmydays @vincent_koc @openclaw love the idea ... what are you doing if the agent decides to ignore the skill? or takes the skill.md as a suggestion and does not follow your steps ?

Gov. - Lineage Claws - CoVerify github.com/SunrisesIllNev… Overall Grade: A (Upgraded from A-) Why the upgrade: I initially missed govern_loop.py, which serves as the operational harness. This is not a prototype — it’s a shippable, production-grade governance framework with: - ReAct-style execution loop - Constitutional enforcement at every step - Durable task memory (progress persistence) - Recovery hooks - Isnad layer integration (signal provenance) - Built-in commitment verification This shifts the evaluation from "promising architecture with missing pieces" to "fully operational governance system with clear roadmap."

The absolute best one hour video to watch this weekend. Karpathy covers engineering’s phase shift, AI psychosis, claws, AutoResearch, a SETI-at-Home-style AI movement, model speciation, jobs data, robotics, MicroGPT, and education.

@vincent_koc @openclaw i developed a plugin guardian ... to block certain hard no no tool execution, e.g. rm -f on ~/clawd or not allowing to restart the gateway (always fun if one agents decides to restart it and all the other agent sessions get killed check it out github.com/coolmanns/open…

@wickedguro What's the repo you are referring to ?

I don't know if people did it, but I cloned the X algorithm repository to my computer and asked Claude Code a simple question: "How Small Accounts Can Go Viral on X - Based on the Algorithm." I know this repo is not super up-to-date with the current algorithm, but here is the output: How Small Accounts Can Go Viral on X - Based on the Algorithm After deep-diving into the actual code, here's what the algorithm reveals: --- The Scoring Formula (Earlybird initial ranking) Score = (replies × 10,000) + (likes × 1,000) + (quotes × 1,000) Replies are weighted 10x more than likes. This is the single most important insight. --- The 5 Easiest Levers for Small Accounts 1. Replies > Everything Else The algorithm values replies at 10x the weight of likes/favorites in the initial Earlybird retrieval score. A tweet with 10 replies scores the same as one with 100 likes. Write tweets that provoke conversation — hot takes, questions, "unpopular opinion" formats. 2. The MTL Normalization Boost (Small Account Advantage) The algorithm has a built-in follower-count normalizer (RescoreMTLNormalization) that explicitly boosts tweets from small accounts. It adjusts scores upward for low-follower authors so they compete fairly against large accounts. This means your engagement-per-follower ratio matters more than raw numbers. 3. Get Just 1 Like From Someone's Follower Network The UTEG (User Tweet Entity Graph) social proof threshold is just 1 favorite (MinNumFavoritedByUserIds = 1). If even one person who follows you likes your tweet, it becomes eligible to appear in the feeds of their followers as an out-of-network recommendation. That single like is the gateway. 4. Topic Alignment via SimClusters The algorithm uses AI embeddings (SimClusters) to match tweets to user interests. The threshold is very low (MinScore = 0.072). If your tweet clearly maps to a topic cluster (e.g., "machine learning", "NBA", "cooking"), it gets surfaced to people interested in that topic — regardless of your follower count. Use clear, topic-specific language rather than vague posts. 5. Bookmarks & Shares Are High-Signal The algorithm tracks bookmarks (PredictedBookmarkScoreFeature) and shares (PredictedShareScoreFeature) as distinct high-value engagement signals. These indicate deeper interest than a casual like. Tweets that are "save-worthy" (threads, tutorials, resources, data) get boosted. --- What Hurts You (picture attached) --- The Viral Loop Path Your tweet gets 1 like from a connected user → Passes UTEG social proof (threshold: 1) → SimClusters matches it to interested users → Shown to ~50-100 new people (out-of-network, 0.75x scale) → MTL normalization BOOSTS your score (small account advantage) → More replies/likes → Earlybird score jumps → Shown to thousands more → exponential growth --- TL;DR — The Easiest Strategy 1. Write reply-bait (questions, debates, hot takes) — replies are 10x more valuable than likes 2. Stay on-topic — clear topic language gets you into SimClusters recommendations 3. Make it save-worthy — bookmarks and shares are premium signals 4. You only need 1 like to break into out-of-network feeds 5. The algorithm literally boosts small accounts via MTL normalization — your low follower count is actually an advantage per-engagement The algorithm is surprisingly favorable to small accounts. The biggest barrier isn't the ranking — it's getting that initial engagement to trigger the viral loop.

Dear @openclaw : - no more rm-f in the workspace - no spinning up docker images without using Komodo API - no more updating your SOUL.md github.com/coolmanns/open…

@VadimAgents @heygurisingh @openclaw I created an issue github.com/openclaw/openc… @openclaw for the missing hooks during tool selection

@CoolmannSa @heygurisingh @openclaw You're right, no native hook — but you can fake it pretty well by having the orchestrator intercept the tool result before passing downstream. Not clean, but it works in production.

🚨BREAKING: Someone just open sourced a complete AI agency and it hit 50K GitHub stars in under two weeks. It's called The Agency. And it's not a prompt template. It's 147 specialized AI agents across 12 divisions -- engineering, design, marketing, product, QA, support, spatial computing, each with its own personality, workflow, and deliverables. Here's what you actually get: → 147 agents across 12 divisions, each with unique voice and expertise → Works natively with Claude Code, GitHub Copilot, Gemini CLI, Cursor, OpenCode, and more → One-command install for any supported tool → Agents have defined missions, success metrics, and production-ready code examples → Full modding support -- build and contribute your own agents → Interactive installer that auto-detects your dev environment → Conversion scripts for every major agentic coding tool → Lua-style Markdown templates with YAML frontmatter Here's the wildest part: Most people use AI like a generalist intern. One model doing everything from writing copy to debugging code. This repo structures AI like an actual company. Specialized roles. Clear responsibilities. Defined workflows between agents. It started as a Reddit thread. Now it has 50K+ stars, 7.5K forks, and contributions from developers around the world. Greg Isenberg called it out. It hit 10K stars in 7 days. This is what the future of AI-assisted development actually looks like. 50K+ GitHub stars. 7.5K forks. 147 agents. 12 divisions. 100% Open Source. MIT License. (Link in the comments)

@VadimAgents @heygurisingh makes sense - unfortunately @openclaw to my knowledge does not have a hook - after the tool Decision

@CoolmannSa @heygurisingh We stopped trying to enforce skills at the prompt level. Instead we built a validation layer — every agent output gets checked before it executes. Dice rolling stops when you add a critic in the loop.

Ep 427: Xandeum Fee Model Explained - STOINC, Retention Fees & Free Market Storage x.com/i/broadcasts/1…

@VadimAgents @heygurisingh How do you manage skill usage ? That’s another big issue I see…. It is very hard to ensure the agent is following skills - it’s like rolling a dice 😝

Exact same path. State engine trap gets everyone. Went back to one orchestrator, tight handoff contracts, Critic agent at zero temp to catch drift before it compounds. Add agents back one at a time only when the previous is stable. Org chart is easy. Memory and context management is where it lives or dies.

@koylanai this is super interesting ... i am not a big believer in skill.md's - they get messed up by agents (at least openclaw) and just because a skill is in a system, doesn't mean that the Agent will be using it.... this is the biggest issues i see with Agents and skills at the moment.

SkillNet is the first paper I've seen that treats agent skills as a network, a three-layer ontology that turns isolated skill files into a structured, composable network. Externalizing knowledge into files isn't enough. You also need to know how those files relate to each other. Layer 1 is a Skill Taxonomy. Ten top-level categories (Development, AIGC, Research, Science, Business, Testing, Productivity, Security, Lifestyle, Other), each broken into fine-grained tags: frontend, python, llm, physics, biology, plotting, debugging. This is the semantic skeleton. It answers "what domain does this skill belong to?" Layer 2 is the Skill Relation Graph. This is where SkillNet diverges from other skill repositories. Tags from Layer 1 get instantiated into specific skill entities (Matplotlib, Playwright, kegg-database, gget). Then four typed relations define how skills connect: > similar_to: two skills do the same thing. Matplotlib and Seaborn both plot. Enables redundancy detection. > belong_to: a skill is a sub-component of a larger workflow. Captures hierarchy and abstraction. > compose_with: two skills chain together. One's output feeds the other's input. This is the relation that enables automatic workflow generation. > depend_on: a skill can't run without a prerequisite. Enables safe execution by resolving the dependency graph before running anything. These four relations form a directed, typed multi-relational graph. Nodes are skills, edges are typed relationships. And the graph is dynamic. As new skills enter the system, LLMs infer relations from their metadata. Layer 3 is the Skill Package Library. Individual skills bundled into deployable packages. A data-science-visualization package contains Matplotlib, Seaborn, Plotly, GeoPandas with their relations pre-configured. You install a package, you get a coherent set of skills that already know how to compose with each other. This is a good example of what comes after a flat package manager. The paper also (you can test here skillnet.openkg.cn) has a science case on a real research workflow: identifying disease-associated genes and candidate therapeutic targets from large-scale biological data. Without encoded relations, the agent figures out the research pipeline from scratch every time. With them, it receives a pre-structured execution plan. The agent still reasons about which genes to focus on and which pathways to investigate. But the pipeline architecture is given. So the skill metadata is actually doing routing work too. The metadata encodes the judgment a domain expert would make when choosing between tools. I also like this framing from the paper: Skills are how memory becomes executable and workflows become flexible. While the network effect and layered architecture is actually useful today, they also acknowledge this: "Low-frequency or highly tacit abilities are difficult to capture, particularly when they resist explicit linguistic description." From my short research career, I'd say the hardest parts are hypothesis generation, experimental design judgment, and interpreting ambiguous results etc. SkillNet handles the structured pipeline well; fetch data → analyze → validate → report. It doesn't handle the creative work where a scientist's (not just in science but in any white-collar field) intuition drives what's worth investigating in the first place. Skills encode "how to run the analysis." They don't encode "what's worth analyzing." That gap is where domain expertise still sits.

@VadimG41837 @heygurisingh sooooo agree with this statement ... did the same ....tried multi agents (5) setup ... went building a state engine and came back to start back at square 1 with one agent ....slowly getting back to adding new agents... but making them flow together is hard ...

@heygurisingh 147 agents is a library. Getting 8 of them to actually hand off context reliably without drift is the hard part. Been running a live pipeline for weeks — the org chart is the easy bit. Memory, state management, and knowing when NOT to spawn an agent is where it gets real.

There's still a few spots left to get free codex Pro subs! developers.openai.com/codex/communit…

96% of clergy given psilocybin at Hopkins called it one of the most spiritual experiences of their lives. Most men left the church. Not the question. adultintraining.us/post/the-churc…

Test post

@classicchins @openclaw @sunilkumar945 @pbteja1998 do you have a repo ? love the dashboard

My @openclaw agents are helping me run 3 businesses and a handful of side projects while I manage everything from a dashboard I built this week with @sunilkumar945 10 agents. Real tasks. I just review and approve. Huge shoutout to @pbteja1998 for showing what was possible. Saw his mission control setup and had to build my own.

@steipete thank you for Lobster Mini Workflow engine! So much better than a simple list of instructions, my #OpenClaw Agent just sometimes ignores ;-) github.com/openclaw/lobst…