Casey Kerr

4 levels of Hermes Agent setup: LEVEL 1: main agent You → Hermes Agent this is your main agent and your prototype area, where you test new workflows and refine them. it doubles as your orchestrator until you have something worth breaking out ---- LEVEL 2: specialized agents You → SEO Agent You → CMO Agent You → Ops Agent once a workflow is solid, break it out into its own agent with its own credentials, memory and scope. --- LEVEL 3: orchestrated team You → Orchestrator ↓ Specialist Agents bring the orchestrator back in. it now steers the company of agents you have built. ---- LEVEL 4: automated team Cron / Events → Orchestrator ↓ Agent Team add task lists so the team works async. cron and events fire jobs, the orchestrator routes them through the task bus, the team handles the work without you ---- take small steps, you DO NOT want to automate slop. if your output at level 1 is mediocre, you are about to scale mediocrity. 20 agents shipping low quality work at speed is worse than 3 shipping great work slowly. I would rather run fewer agents with better output than MAXXING the agent count and spitting out more of the same.

The creator of Claude Code teaches more about vibe-coding in 30 minutes than most tutorials do in hours. Save this — it'll change how you build forever.

I thought I was a Claude Code power user. Then the creator of Claude Code(@bcherny) dropped this thread and I realized I've been using maybe 30% of the tool. >Scheduled automation loops. >Hooks into the agent lifecycle. >Git worktree parallelism. >Fan-out across hundreds of agents at once. >Coding by voice. This thread is mandatory reading if you're building with it.

RIP OpenClaw 💀 Claude now has > Voice mode > Agent Teams > 38+ Connectors > Cowork Projects > Scheduled tasks > Plugin Marketplace > Persistent memory > 1M Context window > Dispatch for remote control > Channels for Telegram & Discord > Claude can use computer to run apps💀

Paper (paper.design) is a new open-alpha AI-native design tool (launched ~early 2026) built on web standards for a single connected canvas. Vs Figma: - Seamless design ↔ code loop (HTML/CSS export, no handoff loss; syncs tokens/styles/components with repos) - AI agents (Claude, Copilot etc.) read/write directly on canvas for layouts, data pulls, boilerplate - Real data/databases from apps/cloud integrated live Figma excels at collab vector design; Paper targets end-to-end "design, share, ship" with less friction. Desktop app available now—early but buzzing as potential disruptor. Check paper.design!

Today, the SDNY prosecutors filed a letter to Judge Failla requesting a retrial date. They want to go again in October. The prosecutors want to retry me on 2 counts the jury couldn't unanimously decide on. A jury of 12 Americans heard 4 weeks of evidence and deadlocked: no verdict on money laundering, and no verdict on sanctions violations. The government's response? Try again to make writing code a crime. @realDonaldTrump declared the "War on Crypto is over." 🇺🇸 AG @DAGToddBlanche's memo: DOJ "is not a digital assets regulator" and won't target mixers for end-user acts. @USTreasury lifted Tornado Cash sanctions entirely. ✅ Also Treasury, March 2026: "Lawful users of digital assets may leverage mixers to enable financial privacy." — official report to Congress under the GENIUS Act. But the SDNY prosecutors — same country, same DOJ — just filed to retry me anyway. 🤔 ⠀ The 2 counts = up to 40 years in federal prison. ⛓️ For writing open-source code. For a protocol I don't control. For transactions I never touched. A jury already couldn't agree this was criminal. But the SDNY prosecutors want to keep trying with the hope of getting a different answer. ⠀ I have a daughter. I have a life in Seattle. I will never stop fighting for freedom. ❤️ But I need to be honest with you: Four weeks of trial. A hung jury. Now they want to do it all over again in October. I have basically exhausted my legal defense funds. And I'm staring down another full federal trial. 😔 Every dollar raised goes directly to keeping this fight alive — attorneys, experts, the full defense apparatus it takes to stand up to the SDNY prosecutors. This isn't abstract. If I can't fund a defense, they win by default. If you care about financial privacy, if you write code and believe that code is speech — this is the moment. 💻🔐 👇

CLAUDE CODE CONECTADO A META ADS 🤯 Hice algo que me sirvió un montón y también les va a servir a ustedes: un mcp para conectarse directamente a Meta Ads y poder chatear con sus datos. Lo mejor de todo es que le sumé la skill que armó Meta en Manus, y tiene toda la info OFICIAL de la plataforma. Le pasan este link a claude code y le piden que los ayude a configurarlo, y listo. Van a tener que crear una app en la consola de desarrolladores de Meta y pasarle un par de tokens y listo. github.com/mathiaschu/met… (BTW, es mi primer repo público 🥹)

@MicahBerkley wow

I had to make a lot of changes this weekend with OpenClaw... but with it being open source... it's been theee most amazing customizable platform in the world to remix for your own personal projects... 🔥 Burned 32M tokens in 48 hours training an AI agent on basic law docs. This weekend I ripped apart my @openclaw gateway config and cut token consumption by ~75% in a single session. Here's what I did on a $7/mo GCP VM (e2-standard-2, 8GB RAM, Debian 12) 👇 🧠 Biggest win: deployed QMD (fully local semantic memory) with 3 GGUF models — embeddinggemma-300M for vectors, qwen3-reranker for relevance scoring, and qmd-query-expansion for smart search (~2.1GB total). Zero API cost for memory retrieval. Previously every message carried the full context payload because QMD had 0 files indexed and 0 vectors embedded. Now it does hybrid BM25+vector search with RRF fusion — targeted retrieval instead of brute force. Also bumped qmd.limits.timeoutMs from 4000ms to 20000ms because cold-starting 3 GGUF models takes ~15s and the default was silently failing. 🤫 ⚙️ Config changes that moved the needle: slashed bootstrapMaxChars from 20,000 to 5,000 (was injecting ~35K tokens per message across 7 files), enabled compaction.memoryFlush so memories persist before context compression, added gemini-2.0-flash as fallback + subagent model for cheap background work, set heartbeat polling to 10s, and capped archiveAfterMinutes at 60. Also caught a silent API quota failure — agent was "completing" runs in exactly 14s with exponential backoff (2s, 4s, 8s) and reporting success with zero output. Direct curl revealed HTTP 429. The agent never told me it was failing. Now I load balanced across 3 keys with 2 levels of failover with other LLMs. 💀 Just dropping this intel, cause I know when I release these projects, the how questions will popup.

CZ wasn’t describing prison. He was teaching HODL technique. 💪

We’re entering Software 3.0: a world where everyone can build personalized software and ship new features in minutes, and the cost is ~$0. Software used to be built by companies with lots of engineers, now it’s built by individuals.

Claude Code isn't for coding. Yeah, I know the name. Ignore it. It's an agent orchestration platform. Coding is just one thing you can point it at. We pointed it at SEO. Built custom agents. Our own workflows. Ran the whole thing through Claude Code.

"Claude Code for the rest of your work" American middle class desk jockeys:

@datarade Good roast

~1% of the population. Please learn math, or we'll need to import more H1Bs.

@datarade I thought you liked going to dinner with me 😥😥😥

I don’t like going to dinners and events with people who draw attention to me and my crew. I prefer to blend in with the walls at all times, I think camouflage is the way. We must make this the norm.

@mminin2010 @CarlHasselskog @bryan_johnson Underrated comment 🤣

@CarlHasselskog @bryan_johnson Dude. With $10B you could clone yourself a whole new body and transplant your brain into it. You might have about 90% chance of not making it through the procedure, but if you do, you'd still have $90B remaining. Expected reward from this gamble is therefore $9B. Pass the dice.

I don’t get the hate for @bryan_johnson. Warren Buffett is 95 and worth $100 billion. Would you trade lives with him? Almost no one would. Which means youth is worth more than $100 billion, and losing at least a billion a year. Ending aging would be the greatest achievement in human history.

@RegenaisanceMan Call the restaurants yourself. Most use msg, put in plastic containers, the list goes on. So tasty. 🤤 but we had to stop eating. 😞

Outside a cheeseburger, pho is the greatest thing to exist. Tripe, lean beef cut, egg, rice noodles, herbs and broth. Elite, could eat this daily.

My conversation with Zcash’s 1000x engineers - Sean Bowe (@ebfull) and Jack Grigg aka Str4d. 0:00 - Project Tachyon 10:20 - Privacy can’t be an afterthought for a protocol 15:24 - Encrypted Bitcoin 22:11 - Soundness bug in original sprout pool 30:15 - Sapling 33:00 - A decade of progress 39:04 - Zcash’s anonymity set 51:29 - Using private money as a passthrough doesn’t work 56:04 - The importance of ledger indistinguishability 58:30 - The ultimate privacy stack 1:05:58 - Tachyon removes all cryptographic bottlenecks of encrypted money 1:08:16 - What’s next after Tachyon? 1:15:25 - Weightless blockchains? 1:18:24 - Ragu 1:19:44 - What the next 12 months look like for Jack and Sean 1:31:22 - “Everything worth doing is worth doing on ambitious timelines.” — Sean Bowe 1:32:16 - Sean’s advice for young engineers 1:34:18 - What got Sean and Jack into Zcash?

Everything is a wrapper A restaurant is a wrapper of a butcher which is a wrapper of a cow which is a wrapper of grass which is a wrapper of photosynthesis which is a wrapper of the sun