NERV Agent coordinates the conversational work. It receives the context, decides between answering, running a tool directly or starting a multi-step task, asks for data when something is missing and delivers results linked to the case.
What the agent knows
- The investigation, the subject and the relevant part of the graph.
- The section you are looking at, your permissions, the enabled plugins and the autonomy level.
- Pending clarifications, identity selections, images and timeline correlations.
- The activity so far: checkpoints, versions and live progress events.
Autonomy levels
| Mode | Behaviour |
|---|---|
| Manual | Every tool or action is confirmed by you. |
| Supervised | Critical or sensitive actions are confirmed by you. |
| Automatic | Standing authorisation within your permissions and policies; the policy layer still applies. |
Autonomy changes how often you are asked, never what is allowed. Even in automatic mode, an action outside your permissions or the active policy does not run.
Tasks and outcomes
Multi-step tasks run in a space of their own for each investigation, with limits and a controlled hand-over of whatever they produce. Their state records the objective, the subject, the plan, the facts found, the blocks met and the products delivered.
Every task ends with an explicit outcome:
| Outcome | Meaning |
|---|---|
| Complete | The objective was met, with evidence behind it. |
| Partial | Some of the objective was met; the rest is stated. |
| Blocked | Something required — a permission, a confirmation, a source — was not available. |
| Failed | The task could not run to the end. |
| No evidence | It ran, but found nothing that supports an answer. |
Governance
Before any action runs, the policy layer checks the tool, the plugin, the investigation, your permissions and the confirmations required. If it cannot decide, the answer is no: the default is to fail closed. A convincing answer from the model never becomes an authorised action on its own.
Good to know
A reply is not proof that something was done. What actually ran, and what it produced, is in the activity of the case; the services, tools and evidence available determine the result.