concept

Multi-Agent Finance Workflows

created 2026-05-04 ai · multi-agent · finance · trading · swarm · dag · patterns

Multi-Agent Finance Workflows

Concrete preset shapes for multi-agent systems in finance — investment, trading, and risk. Pulled from vibe-trading‘s 29 swarm presets as a reference catalog. Useful even outside trading: same shapes apply to any “team of specialists → consensus → sign-off” workflow.

Preset Shapes

Investment Committee

Bull researcher  ─┐
                  ├─→ Risk reviewer ─→ PM (final call)
Bear researcher  ─┘

Pattern: adversarial debate → mediator → decision-maker. Two parallel agents with opposing biases force the mediator to surface real disagreements.

Global Equities Desk

A-share researcher  ─┐
HK/US researcher    ─┼─→ Global strategist
Crypto researcher   ─┘

Pattern: parallel domain specialists → cross-market synthesizer. Each researcher uses domain-specific data sources (e.g. AKShare for A-share, yfinance for HK/US, CCXT for crypto). The strategist’s job is only synthesis — it doesn’t fetch raw data itself.

Crypto Trading Desk

Funding/basis ─┐
Liquidation   ─┼─→ Risk manager
Flow analyst  ─┘

Pattern: feature-specific signal generators → risk gate. Each signal agent has a narrow read on the market; the risk manager decides whether the combined signal clears the trading bar.

Earnings Research Desk

Fundamental analyst ─┐
Estimate revisions  ─┼─→ Earnings strategist
Options pricing     ─┘

Pattern: same as global equities but along the earnings-event axis (fundamentals + sell-side revisions + implied vol).

Macro Rates/FX Desk

Rates analyst       ─┐
FX analyst          ─┼─→ Macro PM
Commodity analyst   ─┘

Quant Strategy Desk

Screener → Factor researcher → Backtest engine → Risk audit

Pattern: sequential pipeline (not parallel debate). Each step has a deterministic-ish output the next step consumes. More like LangGraph Multi-Agent Patterns#1. Handoffs than the parallel-debate shape above.

Technical Analysis Panel

Classic TA ─┐
Ichimoku   ─┤
Harmonic   ─┼─→ Consensus aggregator
Elliott    ─┤
SMC        ─┘

Pattern: ensemble of methodologies → vote/consensus. Useful when there’s no single right answer; the disagreement among methods is the signal.

Risk Committee

Drawdown reviewer ─┐
Tail-risk         ─┼─→ Sign-off
Regime reviewer   ─┘

Architectural Insights

DAG, not chain

All shapes are DAGs (directed acyclic graphs) with parallel-then-merge structure. Pure linear chains (handoff → handoff → handoff) are rare. The merge step is where the real reasoning happens; parallel branches are just specialized data gathering.

Two roles in every team

  1. Specialists — narrow scope, deep tool access, cheap to scale
  2. Synthesizer — broad scope, no raw-data tools, expensive thinking model

The cost-control trick: run specialists on a “sweet spot” tier model (e.g. deepseek-v3.2), reserve the synthesizer slot for a top-tier model (Claude Opus, GPT-5) where reasoning quality compounds.

Adversarial vs ensemble

  • Adversarial (bull/bear): two specialists with mandated opposing priors. Surfaces hidden assumptions.
  • Ensemble (TA panel): N specialists with different methodologies, no mandated priors. Surfaces method disagreement.
  • Pick adversarial when the question has a binary answer; ensemble when it’s continuous (signal strength, regime probability).

Variant: Persona Specialists (instead of Role Specialists)

ai-hedge-fund uses a different shape of specialists: each agent is a famous investor (Buffett, Munger, Cathie Wood, Burry, Taleb…) rather than a functional role (“macro analyst”, “sector analyst”). 14 personas + 4 conventional analysts → risk manager → portfolio manager. Same DAG topology (parallel fan-out → fan-in), different identity model.

The persona shape works when the domain has well-known thought leaders with distinguishable, ideally contradictory positions. See Agent Persona Pattern for when to reach for this vs. role specialists.

Application Beyond Finance

These shapes generalize. For fajb-next:

  • Investment committee → Article publish committee: factual reviewer + style reviewer → editor sign-off
  • TA panel → Source verification panel: archive search + web search + cross-reference → confidence score
  • Quant pipeline → RFP pipeline (in manzas): screen requirements → analyze fit → draft response → review