> ## Documentation Index
> Fetch the complete documentation index at: https://docs.sureops.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# Supervised vs Self-Driving
> How to configure whether sureops drives autonomously or pauses at every stage for your approval.
sureops operates in one of two modes on every active incident: **Self-Driving** (the AI drives end-to-end) or **Supervised** (you control the pace). The mode is a per-incident setting — you can switch between them mid-incident without losing any work.
## The two modes at a glance
| | Self-Driving | Supervised |
| ----------------- | ----------------------------------------------------------------------- | ------------------------------------------------------------ |
| **Who drives** | AI drives all stages | You control the pace; AI is available on demand |
| **Approvals** | Pauses only for action approvals (HITL gates) | Pauses after every stage, waiting for your go-ahead |
| **Override** | Take control at any time | Release to Self-Driving at any time |
| **Best for** | High-volume, lower-severity incidents; teams that have calibrated trust | P1/P2 incidents; new environments; compliance-sensitive work |
| **Notifications** | Real-time notifications when action approval is needed | Notification when each stage completes and awaits input |
***
## Self-Driving mode
In Self-Driving mode, the AI drives the incident through its full lifecycle without waiting for stage-by-stage approval:
```
Detection → Triage → Diagnosis → [HITL gate] → Resolution → Verification → Closure
```
The agent only pauses when an action requires human judgment before it executes. After approval, it continues immediately.
### What you see in Self-Driving mode
* A live timeline as the AI moves through stages
* A notification (Slack, email, or in-app) whenever an approval gate opens
* Full agent reasoning and confidence scores in the incident detail view
* A one-click "Take Control" button that's always available
### Auto-approve configuration
For teams that want fully autonomous operation on lower-severity incidents, Self-Driving mode supports optional **auto-approve** policies. These let specific action types execute without a human approval step.
Auto-approve settings only take effect for **P3 and P4** severity incidents. P1 and P2 incidents always require human approval, regardless of auto-approve configuration. This is a hard safety constraint, not a configurable option.
Two auto-approve knobs are available:
**`autonomous_execution` (recommended)** — Auto-executes recommendations when every proposed action is at or below a risk ceiling you set:
| Setting | Effect |
| ----------------------- | ------------------------------------------------------------------------------ |
| Off (default) | Every recommendation gates on human approval |
| `max_auto_risk: low` | Actions classified as `low` risk auto-execute; `medium` and above still gate |
| `max_auto_risk: medium` | Actions classified as `low` or `medium` risk auto-execute; `high` always gates |
High-risk and destructive actions (rollbacks, config changes, scaling operations classified high-risk) always gate on human approval regardless of the ceiling.
**`auto_approve_hitl`** — The most permissive option. Auto-approves every recommendation regardless of risk level. Intended for fully autonomous low-stakes environments where speed matters most.
To configure, go to **Settings → Environment → AI Behavior → Autonomous execution**.
***
## Supervised mode
In Supervised mode, you hold the wheel. The AI is available as a collaborator at each stage, but it only moves when you say so.
### What you can do in Supervised mode
Run the next stage, then pause again when it completes.
Re-execute the current stage with additional context you provide.
Ask the stage's agent questions. It can query your observability tools in real time and reply with data.
Jump forward to skip a stage, or roll back to revisit a previous one. Requires a reason.
Mark the stage as done with notes — for when you've handled it yourself without agent involvement.
Hand control back to the AI. All context you've provided (notes, chat, re-run results) carries forward into the next agent run.
### Auto-release timer
When you take control, a countdown timer starts (default: 60 minutes). This prevents incidents from being indefinitely parked in Supervised mode with no owner.
* Extend in 15-minute increments (up to 60 minutes per extension) using the **+15 min** button
* A warning appears 5 minutes before expiry
* On expiry, control reverts to Self-Driving and you receive a notification
***
## Taking control mid-incident
You can take control at any time — even while an agent is actively running. sureops uses cooperative cancellation: the agent finishes its current operation (an LLM call, a tool query) and then pauses. It never kills in-flight work.
**What to expect when you take control:**
The UI immediately shows a "handing over..." state. The agent flag is set.
If an LLM call is in-flight, the UI shows "completing current operation." This typically takes 5–30 seconds. The partial findings are preserved.
You see "You have control." All agent findings to this point are visible and usable. The Stage Control panel appears.
Any work the agent completed before the handoff is fully preserved — you never lose analysis, findings, or tool results by taking control.
***
## Configuring the default mode
The default mode is configured at multiple levels. More specific settings override less specific ones.
### Org-level default
**Settings → Organization → AI Behavior → Default mode**
Sets the baseline for all new incidents. All environments inherit this unless overridden.
### Per-environment override
**Settings → Environment → \[environment] → AI Behavior → Mode**
Useful when you want strict supervision in production but full autonomy in staging.
### Per-severity override
**Settings → Environment → \[environment] → AI Behavior → Per-severity modes**
| Severity | Common pattern |
| -------- | --------------------------------------------------------------------- |
| P1 | Always Supervised — high visibility, explicit approval on every stage |
| P2 | Supervised — significant impact warrants oversight |
| P3 | Self-Driving with `max_auto_risk: low` |
| P4 | Self-Driving with auto-approve enabled |
### Per-incident override
The **Agent posture** control on the incident detail page overrides the environment default for that specific incident. The override is visible in the incident header and recorded in the audit trail.
***
## Comparison table
| Scenario | Recommended mode |
| ----------------------------------------------- | ----------------------------------------------------------------------- |
| P1 database outage, production | Supervised — you want full visibility and control over every action |
| P3 high latency on a non-critical service, 2 AM | Self-Driving with `max_auto_risk: low` — let the AI handle it |
| New environment, first month | Supervised — build calibration before trusting autonomous execution |
| High-volume alert storm with known patterns | Self-Driving — agents triage and resolve faster than humans can respond |
| Compliance audit period | Supervised — every action needs an explicit approval record |
***
## Trust progression
sureops is designed for deliberate trust calibration, not automatic escalation. Moving from Supervised to Self-Driving is always an explicit operator decision — the system never promotes itself to more autonomy.
A common progression:
1. **Week 1–2**: All incidents run Supervised. Review agent reasoning, correct misclassifications.
2. **Week 3–4**: Switch P4/P3 incidents to Self-Driving. Monitor audit logs.
3. **Month 2+**: Enable `autonomous_execution` for low-risk actions on P3/P4. Expand risk ceiling only after validating behavior.
4. **Ongoing**: P1/P2 incidents remain Supervised by default for most teams.
The audit trail and incident timeline give you the data to make these decisions — see [Confidence & Audit Trace](/next/operating-modes/confidence-and-audit-trace) for how to read them.