Agents

Agents

MOSAIC runs 25 agents across 4 layers, assembled into one daily cycle by LangGraph.js (mosaic-ts/src/graph/daily_cycle.ts). Each agent lives in mosaic-ts/src/agents/<layer>/; per-layer factories (_factory.ts) and schemas (_schemas.ts) are shared.

State flows layer by layer via per-layer maps in mosaic-ts/src/agents/state.ts (layer1_outputs … layer4_outputs, plus a top-level portfolio_actions mirror).

Layer 1 — Macro (10)

central_bank, china, commodities, dollar, emerging_markets, geopolitical, institutional_flow, news_sentiment, volatility, yield_curve. (_aggregator.ts consolidates Layer-1 output.)

Layer-1 agents call sidecar tools (Tushare/akshare/FRED/Xueqiu/etc.) — e.g. volatility uses get_ivx + get_etf_indicator(510050.SH); emerging_markets uses get_etf_price_data; china uses get_property_data (国房景气指数). The macro layer is 18 tools (see Data Layer for the full list).

Layer 2 — Sector (7)

biotech, consumer, energy, financials, industrials, semiconductor, relationship_mapper. They turn macro context into sector picks → the candidate universe.

The six industry agents share a common core of 8 tools — get_industry_policy, get_xueqiu_heat, get_lhb_ranking, get_broker_research (行业研报), get_etf_holdings, get_stock_data, get_indicators, get_industry_moneyflow. Sector-specific additions: energy also calls get_fred_series; financials also calls get_yield_curve_cn.

relationship_mapper works at the stock level and uses get_lhb_ranking + get_stock_research (个股研报).

Layer 3 — Superinvestor (4)

ackman, aschenbrenner, baker, druckenmiller. Each applies an investing-philosophy lens to the Layer-2 candidate universe (they cite only tickers present in the upstream analysis — never invent codes). All four call get_stock_research (个股研报), get_fundamentals, get_stock_data, and get_indicators. ackman additionally calls get_income_statement / get_cashflow / get_balance_sheet (三张财报); baker calls get_income_statement / get_cashflow; druckenmiller calls get_yield_curve_cn. Each also inherits a few sentiment/policy tools from its base profile (get_industry_policy, get_xueqiu_heat, and/or get_lhb_ranking), so the per-agent lists above are the distinguishing tools, not the complete set.

Layer 4 — Decision (4)

cro, alpha_discovery, autonomous_execution, cio (mosaic-ts/src/agents/decision/). Layer 4 is synthesis-only — no tool calls; each reads upstream state and reasons:

• cro — risk control / veto.
• alpha_discovery — alpha synthesis.
• autonomous_execution — turns decisions into trades (reads CRO + alpha + L3 + Darwinian weights).
• cio — final portfolio. Writes layer4_outputs.cio and mirrors portfolio_actions to the top level (single-writer). The CIO can recommend broad-based ETFs as well as stocks.

MiroFish context injection (opt-in)

When config.mirofish.inject_context is true, the CIO prompt gets an appended MiroFish forward-looking section (latest scenario context, with a "simulation only" disclaimer and an as_of_date anti-lookahead bound). Off by default. See maybeAppendMirofishContext in decision/_factory.ts.

Prompts

Prompts are bilingual and versioned per cohort under prompts/mosaic/ — cohort_default plus the 7 regime cohorts (cohort_bull_2007, cohort_crisis_2008, cohort_bull_2016, cohort_crisis_covid, cohort_recovery_2020, cohort_euphoria_2021, cohort_rate_tightening). Autoresearch evolves these on git branches (see Self-Improvement).

Prompt language follows config.output_language (pickPromptLanguage: Chinese | English | Bilingual).