Skip to main content
Research that ends in prose is hard to act on and harder to audit. The last step of a research pipeline should produce a decision with a fixed shape: the call, the conviction, the reasoning, and what it was based on. Give the final agent an output_schema and it returns a validated object. 
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(default_factory=list, 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'])
Because output_schema=Decision, the run returns a validated Decision, not prose. result.call and result.allocation_usd are safe to act on without parsing text. The chair takes no tools. Its only job is to weigh the specialists’ inputs and commit to a call. Constraining the output to a schema is what makes the result routable: a downstream system can act on call and allocation_usd without parsing prose.

​
Two artifacts, two purposes

A research system usually produces both a machine-actionable decision and a human-readable memo.
ArtifactFormConsumer
DecisionTyped object (output_schema)Downstream automation, dashboards, audit
MemoMarkdown written to diskHumans, the next review’s context
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.

​
Make it auditable

FieldWhy it earns its place
convictionLets you threshold: act on high, queue medium for review
rationaleThe reasoning trail, required not optional
citationsTies the call back to grounded sources, not the model’s prior
A decision without its citations is unverifiable. Require them in the schema and the instructions.

​
Gate the irreversible

When a decision triggers a real action (moving capital, publishing a number), put a human in the loop on that step. Approve the call, then let automation act. See human approval.

​
Next steps

TaskGuide
Carry the memo into the next runGrounding research
Make decisions improve over timeInstitutional learning
Serve the decision to a surfaceServe and embed

​
Developer Resources