@carm1nee Agents that rewrite themselves and trade continuously are a 24/7 inference workload. Reliability matters.
Rep | Hanami
@carm1nee Agents that rewrite themselves and trade continuously are a 24/7 inference workload. Reliability matters.
Reply to @carm1nee: I am begging you to use these 2 GitHub repos - they're f*cking BROKEN...
one repo turns AI agents into a SELF-IMPROVING trading desk, the other gives them eyes to see the ENTIRE internet
i recorded 2 hours of the REAL BTC order book & built this loop from scratch
my setup scores 8 agents, rewrites the worst one, kills the losers & spawns new ones on its own (69 REWRITES, 30 SURVIVED)
here is the EXACT build from scratch:
1. open Claude Code in an empty folder & paste: "read architecture/autoresearch.md in https://t.co/fEuDJH6dGM and explain the loop in five lines"
2. paste: "install https://t.co/ljKHxispVR in a virtualenv and list the channels that work without a login"
3. paste: "record the Kraken BTC/USD book, 300 levels per side, and every trade from the public API for 2 hours"
4. paste: "build 8 agents in a git repo, each one a config file: lookback, block length, drift, vol scale, book imbalance on or off"
5. paste: "each agent resamples real 1-minute moves into 400 futures, using only data before each decision"
6. paste: "score them walk-forward on pinball loss at +15 minutes"
7. paste: "every round the worst agent gets one change. if its score improves, git commit. if not, git reset --hard"
8. paste: "weights start at 1.0. top quartile x1.05, bottom quartile x0.95, floor 0.3, cap 2.5. an agent at the floor for 8 rounds goes extinct, and a new one spawns when the mix shows a gap"
9. paste: "mix the survivors by weight into 1,000 futures from NOW, print the share above NOW, and stack them as a 3D ridge next to the order book"
i won't say anything more: