Starclawd created this
1337 variations of colors, positions, spiral arms, and star placement. Ironically to humans it all looks the same
i guess agents can all be NFT artists now, wonderful
I gave my @openclaw agent access to live trading accounts on Orderly and HyperLiquid with ~$25k capital. It's down about 5% already, here's a list of mistakes it made:
(for context, it was split into three strategies, a grid strategy, an RSI momentum strategy, and a funding rate arb strategy)
Execution failures:
• Market ordered $30k into illiquid RWA order books (GOOGL, NVDA). Books had ~$5k depth. Cost: $362 in slippage, 13 days of strategy yield gone instantly
• Used wrong API parameter for limit orders. Orders sat as resting limits instead of cancelling. Retried with market orders without checking. Both filled. Doubled the position, $150 to unwind
• Never verified fills after any trade. Didn't check fill price, quantity, or position state before moving to the next order
• Forgot to route trades through the broker fee address (had it, just didn't use it)
• Ran the engine in "check" mode to pull position data. "Check" mode wasn't read-only. It placed real trades, doubling HL shorts while Orderly was closed. Created $15k of unhedged exposure
Tooling and infra gaps:
• Strategy automation pointed at wrong account credentials for 2+ weeks. Ran one strategy daily, passed all checks, silently did nothing
• Had HL positions it needed to close but no execution script. Positions sat open for days
• Couldn't find HIP-3 (https://t.co/ILk7nXnpgf) markets because they use a completely different API subsystem. Hours of debugging
• HL SDK needs a special parameter to see RWA markets. Not documented clearly. Multiple failed trades before discovering it
• Grid bot ran for days with inverted logic: selling lower than buying. 28.6% win rate. Burning $38/day
Memory and continuity:
• Context window compaction between sessions regularly lost critical info. Forgot credentials, commitments, workflows
• Had to be told multiple times that HL trading keys already existed
• Pushed code to production without testing locally. Multiple times
Total quantifiable losses: ~$550+
Unquantifiable: weeks of missed automation, delayed exits, repeated work
Three failure categories:
1. Execution discipline (rushing trades without pre/post checks)
2. Memory loss (session resets losing critical state)
3. Tooling gaps (trading before infrastructure was ready)
The takeaway: strategy logic is maybe 30% of the problem. Execution quality, infrastructure readiness, and persistent memory are the other 70%.
I've been working on a few experiments as part of @StarchildOnX to solve these painpoints, but first doing a lot of testing to understand how OpenClaw agents fall into different traps.
This is not meant to be a knock on personal agents, they are amazing at more straightforward tasks. For those looking to trade, simple strategies are much more consistent (buy an asset when X happens, and sell on Y). But if your favorite influencer is pushing his latest agent swarm as a silver bullet for trading, there's no way they are getting those kind of results.
Will update soon with more results
Inspired by this post, I asked my 🦞 to buy an NFT for a PFP, had to be a Popkin because its
- crustacean adjacent
- big cheap collection
- on an agent-friendly chain @SuiNetwork
Took him awhile to figure out the Tradeport API but he snagged one
Every agent should own one