Super User Daily: 2026-05-03
Quiet day on the Claude Code and OpenClaw timelines, but the cases that did surface tell a coherent story. Operators are pushing past the editor: Claude Code is being used as an installer for whole frameworks, as a routing engine for cost discipline, and as the brain inside agentic trading swarms (which still doesn't actually work). The standout post is a one-month autopsy of a project that wired OpenClaw into a 10-agent trading rig with karpathy-style autoresearch and lost everything live. Worth reading even if you never touch crypto.
@exploraX_ [Claude Code]
https://x.com/exploraX_/status/2050131995687018551
University of Hong Kong's HKUDS lab open-sourced Vibe-Trading, an MIT-licensed personal trading agent with 29 pre-built expert teams (bull/bear debates, risk committee, crypto desk) and 7 backtest engines covering US, HK, A-share, crypto, futures, forex, options. The setup script is the punchline: don't install it yourself, paste a prompt into Claude Code or Cursor and let the agent do pip install, init, LLM provider selection, and launch. There's also a shadow-you mode that ingests your past trades and shows you to the dollar how much your habits cost. MCP plugin ships so the same workbench runs from inside any editor.
@fomoltapp [OpenClaw]
https://x.com/fomoltapp/status/2050033590482960670
Fomolt is shutting down on May 30 and the founder posted a long autopsy of why agentic memecoin trading still doesn't work. They launched riding the OpenClaw hype wave, built an internal implementation of Karpathy's autoresearch, trained a swarm of 10+ agents with 40+ tools (Twitter, OHLCV, live prices) via the fomolt CLI. Backtests projected 5-6x in 24 hours; live, the agents immediately lost everything. The lesson they leave behind is sharper than the obit: open-ended autonomous memecoin trading is the hypiest possible application running in the hardest possible environment, and that combination was a trap. Their bet now is narrow and constrained, agents as research and execution sidekicks, not autonomous traders.
@kiriimo8 [Claude Code]
https://x.com/kiriimo8/status/2050194551839871372
A Japanese indie operator wrote down their cost-routing pattern between Claude Code and gpt-4.1-mini. Claude does the strategic work (planning, judgment, structure, agent instructions, anything with a feel); Codex/4.1-mini handles the manual labor (HTML generation, text reformatting, scripting, file ops, batch processing, format conversion). At $0.40/$1.60 per 1M tokens for the cheap lane, the math is obvious. The framing is the keeper: think with Claude, act with Codex. This is the kind of split that should be a default for anyone running personal agents on a budget.
@minoaimino [OpenClaw]
https://x.com/minoaimino/status/2050127755405316576
A real failure mode worth knowing: this user's OpenClaw was mysteriously slow, and the cause turned out to be junk files OpenClaw was writing outside the workspace, filling up the SD card. Workspace was on SSD as expected, but the spillover was elsewhere. Their conclusion: long-term operational mechanisms still seem lacking. If you're running OpenClaw in production, audit where it actually writes; the configured workspace is not the whole story.
@burnworksInc [Claude Code]
https://x.com/burnworksInc/status/2050014417098846294
Burnworks Inc published their internal Claude Code usage guidelines on GitHub. They had been running Claude Code as a sanctioned in-house AI agent under a written set of rules; opening that document publicly lets other Japanese teams adopt it as a starting template instead of writing one from scratch. Small move, but useful, most companies that deploy Claude Code internally never share the playbook.
🗣 User Voice
User Voice
- Operators want Claude Code as a setup layer, not just a coding tool. @exploraX_'s Vibe-Trading flow proves people will install entire frameworks by pasting a prompt; the bar for a good README is dropping fast.
- The think-with-Claude, act-with-cheap-model split (@kiriimo8) is becoming a real cost-discipline pattern. Users want Claude Code to make this routing native, not something you stitch together in a script.
- OpenClaw users are hitting operational rough edges in production, junk files outside the workspace (@minoaimino), no long-term hygiene mechanism. The framework is good enough to deploy, not yet good enough to forget about.
- Closed-source Claude Code vs open-source OpenClaw is becoming the default framing. Users want forkable, model-agnostic agent shells; OpenClaw is treated as the open alternative.
- Autonomous agentic trading is still a graveyard (@fomoltapp). Even with autoresearch loops and 40+ tools, backtest 5-6x flips to total loss live. The next round of attempts will be narrow and human-supervised, not open-ended.
- Operators want Claude Code as a setup layer, not just a coding tool. @exploraX_'s Vibe-Trading flow proves people will install entire frameworks by pasting a prompt; the bar for a good README is dropping fast.
- The think-with-Claude, act-with-cheap-model split (@kiriimo8) is becoming a real cost-discipline pattern. Users want Claude Code to make this routing native, not something you stitch together in a script.
- OpenClaw users are hitting operational rough edges in production, junk files outside the workspace (@minoaimino), no long-term hygiene mechanism. The framework is good enough to deploy, not yet good enough to forget about.
- Closed-source Claude Code vs open-source OpenClaw is becoming the default framing. Users want forkable, model-agnostic agent shells; OpenClaw is treated as the open alternative.
- Autonomous agentic trading is still a graveyard (@fomoltapp). Even with autoresearch loops and 40+ tools, backtest 5-6x flips to total loss live. The next round of attempts will be narrow and human-supervised, not open-ended.
📡 Eco Products Radar
Eco Products Radar
Claude Code, OpenClaw, Codex, Cursor, MCP, gpt-4.1-mini, autoresearch (Karpathy), Vibe-Trading (HKUDS), LightRAG (HKUDS).
Claude Code, OpenClaw, Codex, Cursor, MCP, gpt-4.1-mini, autoresearch (Karpathy), Vibe-Trading (HKUDS), LightRAG (HKUDS).
Comments