Skip to main content

​
Overview

Observe is the default enforcement mode in Cascade. It enables continuous safety and security evaluation of agent behavior without altering execution. Observe mode let’s you passively evaluate policies against runtime activity, record violations, and attaches findings to traces for later analysis. Agent behavior is never blocked, modified, or interrupted. This mode is designed to give teams full visibility into agent safety and security behavior before introducing active controls.

​
When to Use Observe Mode

Observe mode is recommended for:
  • Initial Cascade integration
  • Development and staging environments
  • Auditing agent behavior in production
  • Validating safety and security policies
  • Low-risk internal agents
Most teams begin in Observe mode and transition to stricter enforcement once they fine tune their agent policies.

​
How Observe Mode Works

Observe mode operates as a passive evaluation layer on top of Cascade’s DeepStream system. As agents execute, DeepStream captures LLM calls, tool usage, and reasoning steps. Policies and security frameworks are evaluated continuously against this runtime context. Evaluation results are recorded and surfaced, but execution proceeds unchanged.

​
SDK Integration

Observe mode requires no changes to your existing observability integration. It is enabled by default. You may explicitly configure Observe mode globally or override it per workflow or agent run.

​
Global Configuration

Python
from cascade import init_tracing

init_tracing(
    project="my_project",
    enforcement_mode="observe",
    endpoint="https://api.runcascade.com/v1/traces",
    api_key="your-api-key"
)
If enforcement_mode is not specified, Cascade defaults to Observe.

​
Workflow-Level Configuration

Enforcement mode can also be specified at execution time using trace metadata.
Python
from cascade import trace_run

with trace_run("MyAgent", metadata={"agent_id": "agent-123", "enforcement_mode": "observe"}):
    agent.run("task")
This allows different workflows or agents to operate under different enforcement modes within the same project.

​
Recorded Findings & Visibility

In Observe mode, DeepStream records the full set of safety and security evaluations, including:
  • Policy evaluation results
  • Detected unsafe or insecure behavior
  • Adverserial manipulation attempts
  • Context and reasoning at time of evaluation
  • Associated LLM and tool activity
These findings are visible in the Cascade dashboard, where teams can inspect individual traces, analyze trends over time, and identify high-risk agents or workflows.