274‑agent trading terminal
A thread detailed 'MiroFish'—a Claude‑powered trading terminal made of 274 agents using Bayesian updates and Kelly‑criterion sizing that posted $15k simulated PnL over 14 days. It's a concrete example of agentic orchestration and automated position sizing being tested in simulated trading environments. (x.com)
Attempting to open the X status at returned no accessible content from the public endpoint. (x.com) The MiroFish repository (666ghj/MiroFish) describes a GraphRAG-backed multi‑agent engine that explicitly advertises spawning “thousands” of agents and currently shows about 19.7k stars on GitHub. (github.com) Several outlets report the project was produced rapidly by an undergraduate developer and that Shanda Group founder Chen Tianqiao signed a funding commitment quoted at roughly 30 million RMB in early March 2026. (kucoin.com) The MiroFish docs and DeepWiki break the pipeline into discrete stages—graph building, environment setup, parallel simulation, report generation—and document dynamic graph memory updates and agent profiles that are the same engineering primitives needed to implement per‑agent belief updating. (deepwiki.com) Open‑source Claude‑powered trading projects that mirror the orchestration pattern also exist: degentic‑tools’ Claude‑Code trading terminal, the Clodds Claude‑powered terminal, and several Polymarket trading bots that pair Claude ensembles with Kelly‑criterion sizing. (github.com) A public search of major coverage and the MiroFish repo did not surface an independent, archived performance report that matches a 274‑agent run producing $15,000 simulated PnL over 14 days; that performance figure appears only in the inaccessible X thread linked in the card. (github.com) Operating 274 Claude‑backed subagents in parallel would generate hundreds of LLM tool calls per decision cycle and impose nontrivial state persistence and MCP/Memory engineering requirements described in Claude Code and Anthropic’s agent docs for financial deployments. (code.claude.com)
