Grounding research

Ungrounded research is confident and wrong. A research agent needs to know the rules it operates under, the body of knowledge it can draw on, and what was decided before. Agno layers these so each agent carries the right context without carrying everything.

Layer	Holds	Mechanism
Static context	The mandate, policy, and process shared across queries	Injected into every system prompt
Research library	Company profiles, sector analyses, source documents	RAG over a vector database
Prior work	Past decisions and memos	File navigation tools

Layer 1: static context in the prompt

Rules that apply to every question belong in the prompt, not in retrieval. Load them once and inject them into every agent.

from pathlib import Path

CONTEXT_DIR = Path(__file__).parent / "context"


def load_context() -> str:
    sections = [f.read_text() for f in sorted(CONTEXT_DIR.glob("*.md"))]
    return "\n\n---\n\n".join(sections)


COMMITTEE_CONTEXT = load_context()

instructions = f"""\
You are the Risk Officer on a $10M investment team.

## Committee Rules (ALWAYS FOLLOW)

{COMMITTEE_CONTEXT}

## Your Role
Enforce position limits and sector caps on every recommendation.
"""

Mandate, risk policy, and process are markdown files. The loader places the same text in each agent’s instructions. Model instructions guide behavior; enforce position limits and other hard policy in code or tool permissions.

Layer 2: the research library in RAG

The corpus the agents reason over goes in a shared knowledge base, searched per query. Reuse one configured instance when agents share the same corpus.

from agno.agent import Agent
from agno.models.openai import OpenAIResponses

# Shared instance, imported from a settings module
from agents.settings import team_knowledge

analyst = Agent(
    model=OpenAIResponses(id="gpt-5.5"),
    knowledge=team_knowledge,
    search_knowledge=True,
    instructions="Search the research library before forming a view. Cite what you used.",
)

reply = analyst.run("What does our research say about semiconductor supply?").content
# The instruction tells the analyst to search and cite retrieved material.

search_knowledge is on by default. The agent pulls the relevant profiles and analyses per question instead of carrying the whole library in context.

Layer 3: prior work on disk

Keep complete past memos as files when a reviewer needs the original reasoning trail. Give the reading agent file tools scoped to that archive.

from pathlib import Path

from agno.agent import Agent
from agno.models.openai import OpenAIResponses
from agno.tools.file import FileTools

memos_dir = Path(__file__).parent / "memos"

archivist = Agent(
    model=OpenAIResponses(id="gpt-5.5"),
    tools=[
        FileTools(
            base_dir=memos_dir,
            enable_save_file=False,
            enable_replace_file_chunk=False,
        )
    ],
    instructions="Check prior memos before drawing new conclusions.",
)

The agent that writes new memos gets write access; this one only receives read and search tools. Keep sensitive files outside memos_dir because every file under the tool’s base directory may be readable.

Why three layers, not one

Without	You would lose
The prompt	Instructions and context shared across queries
RAG	A corpus too large to inline
The archive	The reasoning trail behind past decisions

Each layer answers a different question: what are the rules, what do we know, what did we decide.

Next steps

Task	Guide
Make grounded research compound	Institutional learning
End in an auditable artifact	Structured deliverable

Welcome

Get Started

Features

Use Cases

Grounding research

Layer 1: static context in the prompt

Layer 2: the research library in RAG

Layer 3: prior work on disk

Why three layers, not one

Next steps

Developer Resources

​Layer 1: static context in the prompt

​Layer 2: the research library in RAG

​Layer 3: prior work on disk

​Why three layers, not one

​Next steps

​Developer Resources

Layer 1: static context in the prompt

Layer 2: the research library in RAG

Layer 3: prior work on disk

Why three layers, not one

Next steps

Developer Resources