Skip to content

Paper 013 / DeepMind

A Generalist Agent

Gato trains one sequence model on tokenized observations and actions from hundreds of embodied, game, vision, and language tasks.

UPDATED 16 Jul 2026SOURCE-LED REVIEWRESEARCH REVIEW
01

Paper, researchers, and primary source

Major lab research / agents

Gato trains one sequence model on tokenized observations and actions from hundreds of embodied, game, vision, and language tasks.

CONTRIBUTION / 01

Contribution 1

Represented heterogeneous observations and actions in one autoregressive sequence.

CONTRIBUTION / 02

Contribution 2

Trained a single policy across hundreds of tasks.

CONTRIBUTION / 03

Contribution 3

Demonstrated transfer spanning robotics, games, captioning, and dialogue.

02

Research context

agents / 2022

A Generalist Agent places gato inside the broader agents discussion at DeepMind, with generalist agent supplying a second analytical lens. Read together, the source records three advances: Represented heterogeneous observations and actions in one autoregressive sequence; Trained a single policy across hundreds of tasks; and Demonstrated transfer spanning robotics, games, captioning, and dialogue. Keeping those moves together prevents multitask from being detached from its evidence. For an implementation review, the relevant consequence is that A common token interface can unify very different agent skills, although performance and safety still need task-specific evaluation.

03

Methods and evidence reading

1 cataloged method notes

METHOD / 01

Method 1

The experimental design in A Generalist Agent tests represented heterogeneous observations and actions in one autoregressive sequence and trained a single policy across hundreds of tasks against the paper's documented baselines, datasets, model variants, or systems workloads.

How to read the evidence

The evidentiary value of A Generalist Agent comes from the relationship among its reported moves. Two entries define the method-level claim: Represented heterogeneous observations and actions in one autoregressive sequence; and Trained a single policy across hundreds of tasks. The cataloged result is: Demonstrated transfer spanning robotics, games, captioning, and dialogue. On that basis, A common token interface can unify very different agent skills, although performance and safety still need task-specific evaluation. The catalog nevertheless records that the empirical reach of A Generalist Agent stops at architecture choices, compute budget, task distribution, evaluation protocol, documented data, and comparison baselines; broader generalist agent use therefore requires fresh measurements. Reproduction work should separate genuine gato transfer from behavior caused by a changed generalist agent setup. An independent check of A Generalist Agent needs a fixed gato comparison, a declared generalist agent variation, and saved cases where Trained a single policy across hundreds of tasks does not carry over.

04

Findings in the source record

1 paper-specific findings

  1. The reported evidence in A Generalist Agent supports demonstrated transfer spanning robotics, games, captioning, and dialogue.
05

Practical implication for AI builders

DeepMind / 2022

06

Proposed BrokenGPT application

Research blueprint / proposed status

INTEGRATION POINT / 01

Proposed route placement / gato

For a proposed BrokenGPT experiment based on A Generalist Agent, represent tool calls and observations as typed sequence events so BrokenGPT agents can share one orchestration loop without erasing tool-specific permissions. Keep the gato path isolated, versioned, and attributable to this research record.

VALIDATION METRIC / 02

Proposed acceptance test / generalist agent

Validate the proposed gato route against the paper's reported outcome: Demonstrated transfer spanning robotics, games, captioning, and dialogue. The proposed A Generalist Agent test should capture task completion, recovery behavior, trace quality, and intervention rate, with generalist agent error slices reported apart from the headline gato result.

TRADEOFF / 03

Proposed decision boundary / multitask

Balance autonomy, compute, and controllability before promoting the proposed multitask design. Because the paper leaves a different product, another user population, new hardware, changed operating conditions, and a later model revision as open implementation variables rather than consequences established by its experiments, adoption remains conditional on replay under BrokenGPT's selected model, runtime, and policy configuration.

07

Limitations, verification, and source

Boundaries recorded with the paper

Limitations

  • The empirical reach of A Generalist Agent stops at architecture choices, compute budget, task distribution, evaluation protocol, documented data, and comparison baselines; broader generalist agent use therefore requires fresh measurements.
  • The paper leaves a different product, another user population, new hardware, changed operating conditions, and a later model revision as open implementation variables rather than consequences established by its experiments.

PRIMARY SOURCES

  1. 01
    A Generalist Agent

    DeepMind — Primary primary arXiv paper / 12 May 2022 / Scott Reed, Konrad Zolna, Emilio Parisotto, and 17 more

Related research reviews

View all 100 credited research papers

STRAIGHT ANSWERS

Frequently asked questions

01What does A Generalist Agent study?

Gato trains one sequence model on tokenized observations and actions from hundreds of embodied, game, vision, and language tasks.

02Which methods does A Generalist Agent use?

The experimental design in A Generalist Agent tests represented heterogeneous observations and actions in one autoregressive sequence and trained a single policy across hundreds of tasks against the paper's documented baselines, datasets, model variants, or systems workloads.

03What does A Generalist Agent report?

The reported evidence in A Generalist Agent supports demonstrated transfer spanning robotics, games, captioning, and dialogue.

04What is the proposed BrokenGPT application for A Generalist Agent?

Proposed: represent tool calls and observations as typed sequence events so BrokenGPT agents can share one orchestration loop without erasing tool-specific permissions.

MAJOR LAB RESEARCH / PAPER 013

Continue after A Generalist Agent

After A Generalist Agent, browse the full index for adjacent agents research and work from DeepMind.

Open the paper index