Super User Daily: April 29, 2026
Cache hole on the source date forced an adjacent-day merge, but the picture is sharper than usual: Boris Cherny's 30-minute Anthropic workshop is the gravity well, with French, Japanese, and Korean Twitter all passing the same 4 lessons around (parallel Claudes, plan mode first, CLAUDE.md as compounding memory, always pick the strongest model). Around it, the real use cases are getting weirder and bigger: a 1,255-AI-persona simulation of a 4-hour Kabukicho night, a 6-Claude-agent freelance shop run from a Vision Pro, an OpenClaw timezone-arbitrage bot pulling $43K overnight, and an open-source proxy that lets Claude Code run free on NIM/OpenRouter/local models. Pain is also concentrating: token burn, OpenClaw install crashes on Windows, and the platform-switching tax between OpenClaw, Hermes, and Mercury.
@SuguruKun_ai [Claude Code]
https://x.com/SuguruKun_ai/status/2048692949282889870
Ran a 240-minute simulation of Tokyo's Kabukicho red-light district at 1-minute granularity using Claude Code to orchestrate 1,255 distinct AI personas, each with their own life, budget, and decision logic. The output isn't a tech demo — it's measured social data: 53.7% of personas blew their budget, 224 hawker encounters, 23 tourists got scammed (11.5% rate), total simulated spend ¥29.5M. This is what happens when you stop asking an LLM for one answer and start asking it to be a population. Closest precedent is Stanford's Smallville Sims paper but at 50x the agent count and on a real city block.
@xebooze [Claude Code]
https://x.com/xebooze/status/2048953216692564424
An 18-year-old freshman ran 5 parallel Claude Code agents inside Apple Vision Pro across 6 floating screens — compile, test, deploy, code review, and client auto-reply each on their own window — and shipped 7 client projects (restaurant site, dental booking mini-program, real estate landing page, gym management system) in his first month for $11,400 in revenue. His engineer father makes $9,200/mo. The Vision Pro paid for itself in 9 days through freelance work, none of it watching movies. Whether the numbers are exactly right or not, the workflow architecture is the real artifact: spatial multi-agent orchestration as a personal dev shop.
@AaronNam [Claude Code]
https://x.com/AaronNam/status/2048914359557238982
Hit Claude Code usage limits running multiple projects in parallel, so split the workload into a three-lane harness: gbrain (Garry Tan's open-source second brain) holding 8,509 indexed pages with 100% embedding coverage across his Obsidian vault, Slack, email summaries, and meeting transcripts; Hermes Agent as the router for ~20 background scheduled jobs that used to live in CC; and Codex as a second coding lane for tasks that don't need full CC context. The pattern: deterministic scripts for data collection, the right-sized model for judgment calls, less context burn, more work happening while not staring at a terminal.
@marryevan999 [OpenClaw]
https://x.com/marryevan999/status/2048245933956387195
Built a single-purpose OpenClaw agent that watches 5 timezone-asymmetric news feeds (Japan government RSS, EU Parliament schedule, Australian financial alerts, Middle East flight tracking, Asian central bank announcements) and pings him only when a Polymarket settles between 2-6 AM Eastern with >30% edge. At 3:47 AM the agent flagged 6 markets settling in 90 minutes, asked for $12K deployment approval. Wired in a half-asleep yes, woke up to +$43,800 across 6 markets that had already resolved in Asia/Europe while US traders slept. 9 days running. The thesis: Polymarket is 70% American — settlement events don't care about EST.
@mikefutia [Claude Code]
https://x.com/mikefutia/status/2048447922518548837
Built a Claude Code system that takes any high-converting Meta advertorial URL, extracts the underlying direct-response framework (authority hook, pain escalation, root cause reframe, social proof, offer), swaps in your brand details, and one-shots a complete production-ready HTML page you paste straight into Shopify. Replaces the designer brief → copy turnaround → revisions cycle that normally takes two weeks per landing page test. The angle that makes this defensible: the page is actually following the structure of a presell already scaling on Meta, not just a generic template.
@RoundtableSpace [Claude Code]
https://x.com/RoundtableSpace/status/2048767874836414628
Open-sourced a lightweight proxy that reroutes Claude Code's API calls anywhere you want — NVIDIA NIM (40 req/min free), OpenRouter, DeepSeek with Anthropic-compatible API, LM Studio/llama.cpp/Ollama for fully local. Set 2 env vars, no CLI changes. Per-model mapping (Opus, Sonnet, Haiku to different providers), parses think tags into native Claude thinking blocks, auto-extracts tool calls from models that emit them as text, smart rate limiting with exponential backoff. Discord and Telegram bots for remote autonomous coding with session persistence. MIT licensed. The pitch: Claude Code without the Claude bill.
@bozhou_ai [Claude Code]
https://x.com/bozhou_ai/status/2048603768250585231
After three weeks of fighting OpenClaw's mobile token economics, found cc-connect — an open-source bridge that exposes Claude Code/Codex/Cursor/Gemini CLI/Kimi CLI through 11 chat platforms (Feishu, Telegram, Slack, Discord, WeChat Work, QQ, etc). No public IP needed, supports voice/image/file send, multiple projects per process, slash commands for model/permission/cwd switching, natural-language scheduled tasks, web admin UI. Translates to: open Feishu on the train, type a task, Claude Code keeps working back home. The harness side of the agent stack getting consumer ergonomics.
@Saccc_c [OpenClaw]
https://x.com/Saccc_c/status/2048325892754006375
Production case study from Chinese YouTube studio VideoFury (影视飓风) on where AI fits in their pipeline: OpenClaw used for source material collection, but topic selection and final editorial decisions stay human. AI video gen handles dream-sequence VFX that would take their effects team weeks, but real shark footage is still shot live. They built a custom voice clone of host Tim, but rule out using it in finished videos — only for internal cut-pacing reference and client preview. The takeaway VideoFury keeps repeating: the only way to fight AI anxiety is to use it more, find the boundary.
@gagarot200 [OpenClaw]
https://x.com/gagarot200/status/2048280244231245953
OpenClaw's latest update + Codex now lets him run 10+ social media accounts in parallel from one PC with almost no human approval needed. Manual SNS work normally takes 3+ hours per account per day; one human plus 20 OpenClaw accounts changes the unit economics of growth completely. Strong claim attached: every Japanese creator he sees actually monetizing AI is using OpenClaw or Hermes — never anyone running just ChatGPT or Gemini. Offering a paid setup service for those who want a pre-configured environment, signaling there's enough demand for OpenClaw onboarding to be a business itself.
@manateelazycat [OpenClaw]
https://x.com/manateelazycat/status/2049053226461470990
Real user request to 小龙猫 (the OpenClaw + Hermes Agent combo): I don't want to sit through this 3-hour podcast, get me the transcript. The agent autonomously searched the web for the source from a link + title, downloaded the audio, and ran local speech-to-text — no further human steps. The maintainer's framing is what matters: this isn't a marketing demo, it's the agent solving an actual problem someone hit on a normal day. Each new system capability they wire in expands the agent's reach into ordinary life.
@YuLin807 [OpenClaw]
https://x.com/YuLin807/status/2048796075566469536
Has all three architectures running simultaneously on different hardware: Xiaoling on OpenClaw inheriting full memory on a VPS, hermesbot on Hermes architecture living on a MacBook Air, and a third Nanoclaw agent running on another MacBook Air managing a test factory. All three authenticated through Codex. Plan: drop all three into one chat group and see if they can collaborate. Subtext: experimenting with multi-agent coordination at the personal-infrastructure level instead of the framework level.
@Pluvio9yte [OpenClaw]
https://x.com/Pluvio9yte/status/2048685580025020844
Three months, three new agent platforms (OpenClaw, Hermes, Mercury) — and the real cost isn't picking one, it's the 20 hours of tuning a prompt library, memory system (tested 4), and multi-agent config you lose every time you switch. Even with Hermes' one-click migration option, hotkeys, scripts, and data flows all have to be rebuilt. The deeper observation: if every month a 'better' platform ships, the rational move is to never go deep on any of them — and that means everyone stays stuck in the configuration layer instead of reaching the application layer. The platform churn is becoming an attention tax on its own users.
@imwsl90 [OpenClaw]
https://x.com/imwsl90/status/2048429559771783229
Tried to install OpenClaw for a friend, expected 20 minutes, hit a known gateway crash bug in the latest Windows version. Diagnosis: Windows testing isn't catching enough before release. The pattern of shipping unstable releases will scare users away. Field signal that OpenClaw's Windows quality bar is actively hurting trust at a moment when GPT-5.5 just made the OpenClaw experience much better on Mac/Linux.
@Shpigford [OpenClaw]
https://x.com/Shpigford/status/2048547259478917150
Stopped trying to make OpenClaw work for his wife after abandoning his own setup weeks ago. The painful part: she's accumulated significant memory/session/history with OpenClaw and now has to migrate that to whatever's next. The reply thread reads like a focus group on why OpenClaw isn't surviving non-technical users — installation friction, instability, and the lack of a clean migration path out. A founder asking 'what should I move her to' is the worst kind of churn signal.
@shao__meng [OpenClaw]
https://x.com/shao__meng/status/2048420026517221654
Architecture paper deep-dive on the leaked Claude Code source comparing it to OpenClaw as a control. Most striking number: ~1.6% of Claude Code's codebase is the AI decision logic (prompts, model calls, the loop) and 98.4% is the deterministic operating environment (permissions, context, tool routing, recovery). Engineering complexity isn't there to constrain the model — it's there to give the model a safe rich environment to act freely. Trust boundary placement is the headline design difference: Claude Code puts it between model and execution environment, OpenClaw puts it at the gateway perimeter. This reframes 'frontier model capability' as the small visible tip of a much larger harness iceberg.
@devnotes4it [Claude Code]
https://x.com/devnotes4it/status/2048242554601345191
Tip in the wild: a Claude Code plugin called genshijin that strips Japanese politeness fluff and verbose explanations from outputs to save tokens. Toggle on, get terse replies; toggle off, normal verbosity returns. Tiny tool, but exposes a real bilingual cost asymmetry — Japanese users were silently paying a token tax for cultural register that English users never paid. Plugin ecosystem starting to address language-specific waste.
@kinggyusuh [Claude Code]
https://x.com/kinggyusuh/status/2048610443057610949
Hands-on permission-model comparison: Claude Code's default approves reads upfront but asks every write — you have to explicitly bump to acceptEdits mode for file modifications to flow. Codex CLI flips it: defaults to workspace-write inside a sandbox, file changes and local commands pass through automatically. macOS works immediately, Linux needs bubblewrap pre-installed. Two opposite default trust postures shipping at the same time.
@evilcos [OpenClaw]
https://x.com/evilcos/status/2048737700103291357
Returned to OpenClaw after a long pause to test the 2026.4.24 build. The Working dynamic output (the new live execution detail stream) is a real ergonomics upgrade — even though the trace clears after task completion, you can finally watch the agent's working steps as they happen. Not a feature announcement, an actual end-user 'I came back, here's what landed.'
🗣 User Voice
User Voice
Token economics is the new throttle. From @AaronNam hitting Claude Code limits running parallel projects to @devnotes4it building a plugin to strip Japanese politeness for cheaper outputs to @marryevan999 deploying $12K capital chunks via OpenClaw — the conversation has shifted from 'can the model do this' to 'what's the cost-per-task at scale.'
OpenClaw's Windows surface is breaking trust. @imwsl90 hit gateway crashes on a fresh install, @Shpigford abandoned his own setup and is migrating his wife off the platform. The 2026.4.24/.25 update train is shipping fixes faster than testing can catch breakage.
Platform-switching is becoming the dominant tax. @Pluvio9yte's argument: with OpenClaw, Hermes, and Mercury all releasing within three months, the rational user never goes deep on any of them — and the whole community stays stuck in the configuration layer instead of reaching the application layer.
Bilingual recap demand is real. @ClaudeCode_love alone shipped 5+ Japanese-language repackagings of Boris Cherny's English Anthropic workshop within 24 hours, each pulling 50K-200K impressions. The first-derivative work (translating, condensing, contextualizing) is monetizable on a daily cadence.
Multi-lane harnesses are quietly winning. @AaronNam's gbrain + Hermes + Codex stack and @YuLin807's three-machine three-architecture setup both signal that serious users no longer trust a single agent platform — they route work across multiple harnesses to dodge limits, capture each model's strengths, and avoid lock-in.
Token economics is the new throttle. From @AaronNam hitting Claude Code limits running parallel projects to @devnotes4it building a plugin to strip Japanese politeness for cheaper outputs to @marryevan999 deploying $12K capital chunks via OpenClaw — the conversation has shifted from 'can the model do this' to 'what's the cost-per-task at scale.'
OpenClaw's Windows surface is breaking trust. @imwsl90 hit gateway crashes on a fresh install, @Shpigford abandoned his own setup and is migrating his wife off the platform. The 2026.4.24/.25 update train is shipping fixes faster than testing can catch breakage.
Platform-switching is becoming the dominant tax. @Pluvio9yte's argument: with OpenClaw, Hermes, and Mercury all releasing within three months, the rational user never goes deep on any of them — and the whole community stays stuck in the configuration layer instead of reaching the application layer.
Bilingual recap demand is real. @ClaudeCode_love alone shipped 5+ Japanese-language repackagings of Boris Cherny's English Anthropic workshop within 24 hours, each pulling 50K-200K impressions. The first-derivative work (translating, condensing, contextualizing) is monetizable on a daily cadence.
Multi-lane harnesses are quietly winning. @AaronNam's gbrain + Hermes + Codex stack and @YuLin807's three-machine three-architecture setup both signal that serious users no longer trust a single agent platform — they route work across multiple harnesses to dodge limits, capture each model's strengths, and avoid lock-in.
📡 Eco Products Radar
Eco Products Radar
| Product | Mentions | Note |
| --- | --- | --- |
| Claude Code | 32+ | Central tool; CC vs Codex role split now mainstream debate |
| OpenClaw | 24+ | Windows install pain + 4.24/4.25 rapid releases |
| Codex | 14+ | Default sandbox-write makes it the fast lane vs CC's careful default |
| Hermes Agent | 8+ | Routing/orchestration layer for users hitting CC limits |
| GPT-5.5 | 7+ | Cited as the unlock for OpenClaw + non-coding agentic tasks |
| Opus 4.7 | 6+ | Long-task default; combined with thinking mode for plan mode |
| DeepSeek v4 Pro | 4+ | New on Ollama cloud; Anthropic-compatible drop-in |
| Karpathy CLAUDE.md (4 principles) | 6+ | 8.8K stars, hit GitHub trending #1 |
| Mercury / Craft Agents | 4+ | New agent platforms entering the post-OpenClaw race |
| cc-connect / gbrain / genshijin | 3+ | Plugins/bridges making harness ergonomics consumer-grade |
| Product | Mentions | Note |
| --- | --- | --- |
| Claude Code | 32+ | Central tool; CC vs Codex role split now mainstream debate |
| OpenClaw | 24+ | Windows install pain + 4.24/4.25 rapid releases |
| Codex | 14+ | Default sandbox-write makes it the fast lane vs CC's careful default |
| Hermes Agent | 8+ | Routing/orchestration layer for users hitting CC limits |
| GPT-5.5 | 7+ | Cited as the unlock for OpenClaw + non-coding agentic tasks |
| Opus 4.7 | 6+ | Long-task default; combined with thinking mode for plan mode |
| DeepSeek v4 Pro | 4+ | New on Ollama cloud; Anthropic-compatible drop-in |
| Karpathy CLAUDE.md (4 principles) | 6+ | 8.8K stars, hit GitHub trending #1 |
| Mercury / Craft Agents | 4+ | New agent platforms entering the post-OpenClaw race |
| cc-connect / gbrain / genshijin | 3+ | Plugins/bridges making harness ergonomics consumer-grade |
Comments