The Agent Harness Playground: A ReAct-Style Agent Harness with Budgeted Multi-Agent Delegation and Embedded Quantitative Equity Research

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

We describe the Agent Harness Playground, an agentic system built from first principles around a single ReAct-style loop, and extended along three axes usually left implicit in agent frameworks: delegation as a tool, cost as a first-class constraint, and domain research. The harness exposes one primitive: a model call that may emit tool calls. Everything else, including delegation, planning, memory, verification, and quantitative finance, is composed from this primitive. Multi-agent behaviour is not an orchestration layer above the loop. Instead, it is a nested invocation of the same loop reached through an ordinary tool call. This makes delegation visible in the trace, bounded by one shared budget, and typed at the hand-off boundary. The system bounds its own context and bill with per-observation caps, tool-result memoisation, prompt-cache breakpoints, compaction, and three simultaneous ceilings on steps, dollars, and tokens. Finally, the harness embeds a quantitative equity-research subsystem with a native factor/indicator engine and a subprocess bridge to Microsoft Qlib. The subsystem computes the Alpha158/Alpha360 feature sets, supports an arbitrary-expression language, cross-sectional learned models including LightGBM, Ridge, XGBoost, and CatBoost, an out-of-sample simulator, and a signal-analysis report covering Spearman information coefficient, ICIR, quantile forward returns, and long-short spreads. We give the algorithms and their mathematics, report measured figures from the implementation's own traces, and compare the design to prior work.