Investment Committee

Seven specialists across four team modes, sharing one knowledge base and one learning store.

The Investment team is the most ambitious multi-agent demo in the repo. Seven specialist agents collaborate across four distinct team configurations to evaluate stocks, weigh risk, and produce investment memos. All seven share one knowledge base and one learnings store.

Agent	Role	Tools
Committee Chair	Final decision-maker, capital allocator	none directly
Financial Analyst	Fundamental valuation, balance sheet, earnings	YFinanceTools
Market Analyst	Macro context, sector trends, breaking news	YFinanceTools, Exa MCP, optional ParallelTools
Technical Analyst	Price action, momentum indicators, entry/exit timing	YFinanceTools
Risk Officer	Downside scenarios, position sizing, mandate compliance	YFinanceTools
Knowledge Agent	Research RAG and memo archive navigation	FileTools (read-only)
Memo Writer	Synthesizes analyst inputs into formal memos	FileTools

Why a team for this

The single-agent failure mode for investment analysis is bullish bias: one model talks itself into a thesis. A multi-agent committee structurally avoids this. The Risk Officer is required to argue against, the Technical Analyst is required to look at chart-level evidence, and the Chair has to weigh both.

This is “agents as roles” rather than “agents as task fanout”. Each member has a distinct charter that the Chair must consider.

Four team modes, one set of agents

The demo ships four distinct teams built from the same agent set. Each demonstrates a different multi-agent architecture:

Team	Mode	Right when	Try
Investment Coordinate	TeamMode.coordinate	Open-ended question that benefits from multi-analyst discussion	”Should we invest $2M in NVIDIA?”
Investment Route	TeamMode.route	Single-specialist question that should land with exactly one agent	”What’s AAPL’s P/E ratio?” (routes to Financial Analyst)
Investment Broadcast	TeamMode.broadcast	High-stakes decision where you want independent analyst opinions in parallel	”Full committee review: evaluate Tesla for $2M.”
Investment Tasks	TeamMode.tasks	Multi-step task that needs decomposition and parallel work	”Deploy $10M across the top 5 AI stocks, no position over 30%.”

Pick the team in the AgentOS UI based on the shape of the request. See Team modes for the underlying mechanics.

Three-layer knowledge

Every agent on the team draws from the same three-layer context:

Layer	What it holds	How it’s accessed
Static context	Fund mandate, risk policy, evaluation process	Injected into every agent’s system prompt
Research library	Company profiles, sector analyses	PgVector hybrid search (RAG)
Memo archive	Past investment memos and decisions	FileTools over the memos directory

Same “context is everything” pattern Dash uses, applied to financial decision-making instead of SQL.

Shared learnings

Every agent writes to one investment_learnings knowledge store with namespace="global":

from agno.learn import LearnedKnowledgeConfig, LearningMachine, LearningMode

investment_learnings = create_knowledge("Investment Learnings", "investment_learnings")

_learning = LearningMachine(
    knowledge=investment_learnings,
    learned_knowledge=LearnedKnowledgeConfig(
        mode=LearningMode.AGENTIC,
        namespace="global",
    ),
)

After a few decisions the store accumulates patterns like:

• “Tech stocks with negative free cash flow get downgraded”
• “Avoid recommending positions during earnings week”
• “Risk Officer’s concerns weighted 1.5x for crypto exposure”

The Chair retrieves these on every new evaluation. The committee gets sharper over time without anyone manually building the framework.

File output

The Memo Writer uses FileTools to produce a markdown memo per evaluation:

teams/investment/memos/
├── 2026-04-22_TSLA.md
├── 2026-04-23_NVDA.md
└── 2026-04-24_META.md

Each memo captures the committee’s full reasoning, the dissents, and the final recommendation. Useful for retrospective analysis: which calls were right, which were wrong, and what the committee learned from each.

Mixing models

The demo runs every agent on the same default model (gpt-5.4 via OpenAIResponses). For real workloads, mixing models per role is a one-line change:

from agno.models.anthropic import Claude
from agno.models.google import Gemini

committee_chair = Agent(
    model=Claude(id="claude-opus-4-6"),   # Strong synthesis for the final decision
    ...
)

market_analyst = Agent(
    model=Gemini(id="gemini-2.5-pro"),    # Long context for filings
    ...
)

The team coordinates regardless of which model each member uses. The standalone investment-committee repo runs Claude Sonnet 4.6 across the analysts and Claude Opus 4.6 on the Chair, which is the production-grade configuration.

See it in action

Try one prompt against each team:

# Route: single specialist
@InvestmentRoute What's NVDA's current trend?

# Coordinate: open-ended discussion
@InvestmentCoordinate Should we invest $2M in NVIDIA?

# Broadcast: parallel independent opinions
@InvestmentBroadcast Full committee review: evaluate Tesla for $2M.

# Tasks: multi-step decomposition
@InvestmentTasks Deploy $10M across the top 5 AI stocks, no position over 30%.

Each prompt triggers the right shape of work. Expect 30-60 seconds for the parallel analyst teams, faster for route, longer for tasks.

Source: teams/investment/. For a standalone version with Claude across the team and a deterministic 5-step pipeline workflow, see the investment-committee repo.