Skip to main content
The Injector agent shows how to give an agent access to runtime configuration without baking it into the prompt or the tool definitions. 
from agno.agent import Agent

DEPENDENCIES = {
    "config": {
        "app_name": "AgentOS Demo",
        "version": "2.1.0",
        "region": "us-east-1",
    },
    "feature_flags": {
        "dark_mode": True,
        "beta_features": False,
        "ai_suggestions": True,
    },
    "user_preferences": {
        "theme": "dark",
        "timezone": "America/New_York",
    },
}

injector = Agent(
    id="injector",
    name="Injector",
    model=MODEL,
    db=agent_db,
    tools=[get_config, check_feature_flag, get_user_preference],
    dependencies=DEPENDENCIES,
    add_dependencies_to_context=True,
)
add_dependencies_to_context=True makes the dependencies visible in the system prompt. The agent can refer to them directly:
“Is dark mode enabled?” → reads feature_flags.dark_mode from dependencies, answers “Yes.”
Tools can also access dependencies via RunContext: 
from agno.run import RunContext
from agno.tools import tool

@tool
def get_config(run_context: RunContext, key: str) -> str:
    """Read a config value injected from dependencies."""
    return str(run_context.dependencies["config"].get(key, "not set"))
The RunContext parameter is auto-injected. Tools that take it get the full request scope: dependencies, user_id, session_id, metadata, messages, even the parent agent.

​
Static vs per-request dependencies

The Injector example uses static dependencies (defined at agent construction). For per-request dependencies, set them on the run: 
result = agent.run(
    "What region am I in?",
    dependencies={"region": "eu-west-1"},     # request-scoped
    user_id="user-123",
)
Per-request dependencies override the agent-level ones for that single run. Useful for multi-tenant apps where each request brings its own tenant ID, feature set, or DB connection.

​
What belongs in dependencies vs memory vs knowledge

Lives inWhen
dependenciesPer-request runtime values: feature flags, tenant config, request scope
MemoryUser-level facts that should survive sessions
KnowledgeContent the agent should be able to retrieve and cite
session_stateMutable state for one session only (see Taskboard)
The lines blur, but the rule of thumb is that dependencies are read-only and ephemeral, memory and knowledge are persistent, and session state is mutable but session-scoped.

​
See it in action

@Injector is dark mode enabled?
@Injector what's the current app version and region?
@Injector what's my preferred timezone?
@Injector list all enabled feature flags
Source: agents/injector/agent.py

​
Next

Session State →