Skip to main content

​
DeepAgents

DeepAgents is B-Bot Hub’s mode for work that benefits from planning and intermediate artifacts. A DeepAgent can use a task list and workspace files while working through a multi-step request. The exact tools and actions available depend on the Agent, session, and environment.

​
Core Concepts

Tasks

A visible plan can divide a goal into smaller steps and show progress.

Workspace Files

Inputs, notes, code, data, and outputs can be organized in the workspace.

Tools

Configured tools may let the DeepAgent retrieve information, process files, run code, or interact with connected services.

Human Review

Human-in-the-loop controls can require review before selected actions.

​
Conceptual Workflow

A task marked complete only indicates workflow state. It does not prove that the result is correct.

​
When to Use DeepAgents

DeepAgents can be useful for:
  • Projects with several dependent steps
  • Work that produces or revises multiple files
  • Analysis that needs visible assumptions and intermediate checks
  • Long requests that benefit from a maintained task list
  • Workflows where people need to review proposed actions
Standard chat may be simpler for a direct question, a short rewrite, or another task that does not need a workspace.

​
Giving a DeepAgent a Good Request

Include:
  • The required outcome and audience
  • Inputs and source-of-truth material
  • Constraints, exclusions, and quality criteria
  • Allowed tools and actions
  • Required file names or output formats
  • Actions that require approval
  • A definition of done
Example:

​
Human-in-the-Loop

Use human review for actions that are sensitive, irreversible, externally visible, or costly. Examples include:
  • Sending messages or publishing content
  • Changing records or permissions
  • Running unfamiliar code
  • Using private data with an external service
  • Triggering purchases, payments, or deployments
Human-in-the-loop reduces risk but does not make an action safe by itself. The reviewer still needs enough context to understand the proposal.

​
Workspace Practices

  • Keep original inputs separate from generated files.
  • Use descriptive file names.
  • Record assumptions and decisions.
  • Review generated code before execution.
  • Download or retain important outputs according to your data policy.
  • Verify workspace behavior and retention for the specific workflow.
Grant only the tools and permissions needed for the task. A system prompt is not an access-control boundary.

DeepAgents Workspace

Learn how to enable and use workspace features.

Tasks

Learn about scheduled and on-demand tasks.

Data Analysis

Apply DeepAgents to a reviewable analysis workflow.

Content Creation

Organize a multi-stage editorial workflow.