> ## Documentation Index
> Fetch the complete documentation index at: https://docs.cascadeflow.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Harness Modes

> Three harness modes — off, observe, and enforce — with rollout guidance for production deployments.

cascadeflow operates in one of three modes, set at initialization.

## Modes

### `off`

No tracking, no enforcement. The harness is completely disabled. This is the default.

```python theme={null}
cascadeflow.init(mode="off")
```

### `observe`

Track all metrics and decisions, but never block execution. Every LLM call and tool execution is recorded with full decision traces. Actions are computed but not enforced — `applied` is always `false` in trace records.

<CodeGroup>
  ```python Python theme={null}
  cascadeflow.init(mode="observe")
  ```

  ```typescript TypeScript theme={null}
  import { CascadeAgent } from '@cascadeflow/core';

  // TypeScript uses CascadeAgent with quality config for observe-like behavior
  const agent = new CascadeAgent({
    models: [
      { name: 'gpt-4o-mini', provider: 'openai', cost: 0.000375 },
      { name: 'gpt-4o', provider: 'openai', cost: 0.00625 },
    ],
  });
  ```
</CodeGroup>

Use `observe` for:

* Initial production rollout to validate metrics before enforcing
* Shadow-mode testing to understand what the harness would do
* Cost and usage analytics without affecting agent behavior

### `enforce`

Track all metrics and enforce constraints. When a hard cap is hit (budget, tool calls, latency, energy) or a compliance violation is detected, the harness takes action: `stop`, `deny_tool`, or `switch_model`.

```python theme={null}
cascadeflow.init(mode="enforce")
```

Use `enforce` when:

* You have validated metrics in `observe` mode
* You need hard budget caps to prevent runaway costs
* Compliance requirements mandate model gating

## Rollout Guidance

Recommended rollout sequence for production:

1. **Deploy with `observe`** — No risk to agent behavior. Collect metrics, review decision traces, validate that the harness sees what you expect.

2. **Review traces** — Check that compliance allowlists, budget calculations, and KPI scoring match your expectations.

3. **Switch to `enforce`** — Once validated, change the mode. The harness will now enforce constraints.

4. **Monitor** — Use `session.summary()` and `session.trace()` to monitor enforcement in production.

```python theme={null}
import os

# Environment-driven mode selection
mode = os.getenv("CASCADEFLOW_MODE", "observe")
cascadeflow.init(mode=mode)
```

## Mode Behavior Matrix

| Behavior            | `off` | `observe`                 | `enforce`                |
| ------------------- | ----- | ------------------------- | ------------------------ |
| Cost tracking       | No    | Yes                       | Yes                      |
| Latency tracking    | No    | Yes                       | Yes                      |
| Energy tracking     | No    | Yes                       | Yes                      |
| Decision traces     | No    | Yes                       | Yes                      |
| Budget enforcement  | No    | No                        | Yes                      |
| Tool call gating    | No    | No                        | Yes                      |
| Compliance gating   | No    | No                        | Yes                      |
| `session.summary()` | Empty | Full metrics              | Full metrics             |
| `session.trace()`   | Empty | Decisions (applied=false) | Decisions (applied=true) |

<Tip>
  **Rollout guide:** [Getting Started: Rollout Guide](/get-started/rollout-guide) | **GitHub:** [examples/production\_patterns.py](https://github.com/lemony-ai/cascadeflow/blob/main/examples/production_patterns.py)
</Tip>