Loop Daily: 2026-05-03
Light volume on the autoresearch and agentic-loop tag today, but one post does the heavy lifting. The fomolt team published a one-month autopsy of running an internal Karpathy-style autoresearch loop against a 10-agent memecoin trading swarm with 40+ tools, and the receipts are brutal: backtest projected 5-6x in 24 hours, live trading lost everything. Their conclusion echoes what people running real autoresearch have been saying quietly: the loop works, the environment was wrong. Open-ended autonomy in adversarial markets is still a graveyard; narrow, scoped sidekick loops are where the next year of progress will land.
#1
@fomoltapp
https://x.com/fomoltapp/status/2050033590482960670
The cleanest write-up of an autoresearch loop failing in production this year. They built an internal implementation of Karpathy's autoresearch and pointed it at trading. Ten agents in a swarm, 40+ tools wired in via a custom CLI (Twitter, OHLCV, live prices), scoped sessions where the swarm could iterate freely on strategies. The loop did its job: agents kept exploring, kept reporting, kept improving against the eval. Backtest said 5-6x daily. Live trading lost everything inside an hour. The takeaway they leave behind, and what every loop builder should pin to the wall: 'open-ended autonomous memecoin trading is the hypiest possible application running in the hardest possible environment, and that combination was a trap.' The next iteration: agents as research and execution sidekicks for humans, or agents in tightly scoped domains. Not autonomous traders.
https://x.com/fomoltapp/status/2050033590482960670
The cleanest write-up of an autoresearch loop failing in production this year. They built an internal implementation of Karpathy's autoresearch and pointed it at trading. Ten agents in a swarm, 40+ tools wired in via a custom CLI (Twitter, OHLCV, live prices), scoped sessions where the swarm could iterate freely on strategies. The loop did its job: agents kept exploring, kept reporting, kept improving against the eval. Backtest said 5-6x daily. Live trading lost everything inside an hour. The takeaway they leave behind, and what every loop builder should pin to the wall: 'open-ended autonomous memecoin trading is the hypiest possible application running in the hardest possible environment, and that combination was a trap.' The next iteration: agents as research and execution sidekicks for humans, or agents in tightly scoped domains. Not autonomous traders.
#2
@LabileTosky66
https://x.com/LabileTosky66/status/2050091583555371016
An autonomous loop with a hard exit clause: their Arc Agent now takes an order, locks payment in a contract, and runs to fulfill. If the agent doesn't execute within 24 hours, the buyer is automatically refunded; if it does, the seller is paid. Small architecturally, but it's the right shape β agentic loops that touch money need a deterministic timeout the user can rely on, not just a hopeful retry policy. Most of the 'autonomous agent' demos hand-wave this part.
https://x.com/LabileTosky66/status/2050091583555371016
An autonomous loop with a hard exit clause: their Arc Agent now takes an order, locks payment in a contract, and runs to fulfill. If the agent doesn't execute within 24 hours, the buyer is automatically refunded; if it does, the seller is paid. Small architecturally, but it's the right shape β agentic loops that touch money need a deterministic timeout the user can rely on, not just a hopeful retry policy. Most of the 'autonomous agent' demos hand-wave this part.
π‘ Eco Products Radar
Eco Products Radar
autoresearch (Karpathy), OpenClaw, fomolt CLI.
autoresearch (Karpathy), OpenClaw, fomolt CLI.
Comments