---
id: "framework-deepmind-autonomy-levels"
type: "framework"
source_timestamps: ["00:24:27"]
tags: ["autonomy", "taxonomy"]
related: ["entity-google-deepmind", "concept-intent-engineering"]
steps: ["Observer", "Consultant", "Collaborator", "Approver", "Operator"]
sources: ["s24-prompt-engineering-dead"]
sourceVaultSlug: "s24-prompt-engineering-dead"
originDay: 24
---
# Google DeepMind's 5 Levels of AI Agent Autonomy

## Overview

A taxonomy attributed by the speaker to researchers at [[entity-google-deepmind]], categorizing AI agents by the level of autonomy and human oversight they require. Used in this source to argue that **higher autonomy demands more rigorous [[concept-intent-engineering]]** — you cannot let a system *act* without first encoding what it should *want*.

## The Five Levels

| Level | Name | Behavior | Human Role |
|---|---|---|---|
| 1 | **Observer** | Watches and reports | Full control |
| 2 | **Consultant** | Provides advice or drafts | Takes all action |
| 3 | **Collaborator** | Works alongside iteratively | Co-creates |
| 4 | **Approver** | Acts, but human signs off pre-execution | Gatekeeper |
| 5 | **Operator** | Operates entirely autonomously | None / oversight only |

## Implication for Intent Engineering

- **Observer / Consultant**: minimal intent encoding required — the human is the alignment safety net.
- **Collaborator**: requires lightweight intent hints (preferences, examples).
- **Approver**: requires explicit decision criteria so the human review is meaningful, not rubber-stamp.
- **Operator**: requires *fully* machine-readable intent — every tradeoff, every escalation rule, every boundary. This is where [[concept-machine-readable-okrs]] and the [[framework-intent-gap-layers|three-layer stack]] become non-negotiable.

## Klarna in the Levels Model

Klarna's customer service agent ([[claim-klarna-intent-failure]]) was effectively an **Operator** — fully autonomous, no human in the loop on most contacts — deployed without the intent infrastructure required for that level. The framework predicts exactly the failure mode that occurred.

## Enrichment Caveat

The enrichment overlay was **unable to verify** a specific Google DeepMind paper proposing these exact five levels with these exact names. Related autonomy taxonomies exist across the industry (OpenAI's levels, SAE-style frameworks). Treat the *attribution* as speaker-asserted while the *taxonomy itself* remains a useful conceptual tool.



## Related across days
- [[framework-5-levels-vibe-coding]]
- [[concept-intent-engineering]]
- [[claim-klarna-intent-failure]]
- [[framework-intent-gap-layers]]