SAFE-Agent: An Internal-Market View of Neuro-Symbolic Trading Agents — Capital Allocation, Concentration, and Ecosystem-Level Feedback

An AI agent that reallocates capital across learned policies operates an internal mar ket: realized rewards determine time-varying budget shares for competing strategy providers. We study this mechanism through SAFE-Agent, a neuro-symbolic trad ing system whose LLM-seeded, genetically refined Hybrid Symbolic Expressions combine predictive signals with explicit risk guards and execute in a typed OCaml interpreter. Interpreting its Hedge controller as a no-regret allocation rule exposes an economic trade-off. The same exponential update that limits hindsight regret relative to a fixed expert pool also concentrates capital, which we characterize with a Herfindahl–Hirschman index (HHI) and compare with logged allocation dynamics. A stylized multi-agent model then connects this micro mechanism to ecosystem risk: when independently operated agents share data, grammars, and threshold guards, correlated de-risking can amplify volatility through performative feedback. Empirically, a point-in-time S&P 500 study uses all 635 securities repre sented by dated membership intervals rather than a fixed ex-post snapshot. Over 2020–2024, SAFE-Agent records Sharpe 2.387 and 6.04% maximum drawdown under approximately 0.90 gross exposure, 8.8% one-way daily turnover, 5 bps per-side costs, and an 18% name cap. Baseline, ablation, cost, concentration, and look-ahead diagnostics show how typed search and adaptive allocation improve the return–drawdown trade-off while exposing the accompanying rise in concentration. Together, the formal analysis and mechanism-level evidence establish a cross-scale framework for auditing the efficiency, concentration, and feedback consequences of learned allocation systems.