Governed Self-Evolution for Autonomous Investing: A Multi-Agent LLM System with an Invariant Core
4.20T1 sourcearXiv cs.MA
Source record
Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2610.30557.
Pipeline notes
The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.
SummarySteward is a multi-agent LLM system for autonomous investing: nine role-specialized agents (analysts, portfolio manager, risk manager), a seven-layer shared memory, and a self-evolution loop in which a reflector proposes prompt learnings that a deterministic gate applies reversibly around an invariant core of order arithmetic and risk rules. The paper reports a 102-day live deployment (1,632 LLM calls), replay comparisons, and a 2015–2026 backtest.
Why it mattersThe invariant-core-plus-reflector pattern is a tested approach to bounded self-evolution in agent systems, with live deployment numbers and replay evidence that go beyond typical agent demos.
Cited by
No citations on record.