Structured deliverable

Make the final pipeline step return a typed decision that downstream code can validate and use. Define the call, conviction, allocation, rationale, and citations in a Pydantic schema, then pass it as the final agent's output_schema.

These are composition examples adapted from the Investment Team application. Follow its clone, database, credentials, and research-loading setup for a complete application. Its agents/, teams/, and workflows/ define the components referenced here; agents/settings.py supplies shared knowledge and paths.

Import those components into your application module, or define your own before composing them. The application uses Gemini; the OpenAI variants on these pages require uv pip install "agno[openai]" and OPENAI_API_KEY as well. Shared knowledge/storage still needs the application's database and embedding-provider setup. Prose context, analyst-output variables, and local archives must be supplied by your application.

from typing import List, Literal

from agno.agent import Agent
from agno.models.openai import OpenAIResponses
from pydantic import BaseModel, Field


class Decision(BaseModel):
    call: Literal["BUY", "HOLD", "PASS"] = Field(..., description="The committee decision")
    conviction: Literal["low", "medium", "high"] = Field(..., description="Confidence in the call")
    allocation_usd: float = Field(..., description="Dollar allocation, 0 if not BUY")
    rationale: str = Field(..., description="Why, referencing the analyst inputs")
    citations: List[str] = Field(..., description="Sources and prior memos used")


chair = Agent(
    name="Committee Chair",
    model=OpenAIResponses(id="gpt-5.5"),
    output_schema=Decision,
    instructions=(
        "Synthesize the analyst inputs into a decision. Every BUY needs a "
        "dollar amount. Every decision must reference at least one risk."
    ),
)


def briefing(*analyst_outputs: str) -> str:
    return "Analyst inputs:\n\n" + "\n\n".join(analyst_outputs)


result = chair.run(briefing(market, fundamentals, technicals, risk)).content
# Decision(call='BUY', conviction='high', allocation_usd=2_000_000.0,
#          rationale='Momentum and fundamentals align; sized within the
#                      sector cap the Risk Officer set.',
#          citations=['memo:NVDA-2024Q3', 'research:semiconductors'])

output_schema=Decision requests a parsed Decision. If output parsing fails, content can remain a string. Check that the run completed and isinstance(run.content, Decision) before reading fields or acting on the result.

The chair weighs the specialists' inputs and commits to a call. Its configuration omits tools and supplies a briefing as context. That does not guarantee that every conclusion is supported by the briefing; validate claims against the supplied evidence.

Decision and memo

A research system usually produces both a machine-actionable decision and a human-readable memo.

The memo is written by a dedicated agent with file tools and a fixed template, then archived. The next review reads it back as prior work. The decision is the row you store and act on.

Required decision fields

The schema requires a citations field but allows an empty list and does not check source identifiers. Before a downstream action, application code should require the necessary citations, resolve them against the actual research, and enforce its allocation and action thresholds.

Add approval for consequential actions

When a decision triggers a real action, such as moving capital or publishing a number, add human approval before the action executes. See human approval.

Next steps

Developer Resources

• Structured output
• Workflows cookbook
• Human approval