Skip to main content

​
Agent Architecture

B-Bot Hub separates the definition of an AI Agent from the experience used to reach it. This page describes that relationship at a conceptual level. It is not a specification of internal services or execution order.

​
Main Layers

User Experience

A chat or another Distribution Channel receives input and presents the result.

Agent Configuration

Identity, system instructions, abilities, model choices, and permitted capabilities define the Agent.

Context and Capabilities

Conversation context, knowledge, apps, tools, modalities, and team members provide task-specific inputs or actions.

Model

The selected model interprets the available context and generates a response or proposes tool use.

​
Conceptual Request Flow

Not every request uses every layer. A direct text question may need only instructions and a model. A multi-step DeepAgents request may also use tasks, workspace files, and review points.

​
Configuration and Authorization

Instructions tell the model how it should behave. Permissions determine what the surrounding system allows it to access or do. These are different controls.
  • Use the system prompt for role, priorities, and behavioral boundaries.
  • Use app, tool, and account permissions for access control.
  • Use human-in-the-loop review for sensitive proposed actions.
  • Use Distribution Channel settings for how an experience is exposed.
Do not treat a prompt instruction such as “never access private data” as a substitute for technical permissions. Remove access that the workflow does not need.

​
State and Context

The context available to a request can include the current conversation, Agent instructions, retrieved knowledge, tool results, and workspace artifacts. Availability and retention depend on the feature and configuration. Because context is limited and can change:
  • Keep instructions concise and resolve conflicts.
  • Retrieve only relevant source material.
  • Store large intermediate artifacts in workspace files when appropriate.
  • Confirm that important state is present before continuing a task.
  • Do not assume that information from another session or channel is available.

​
Failure Boundaries

Design each layer to handle failure clearly:
  • Invalid or unsupported input should produce an actionable message.
  • Missing knowledge should produce uncertainty or a request for information.
  • Tool failure should not be described as success.
  • Sensitive actions should stop at the configured review point.
  • The Distribution Channel should communicate the final status to the user.

AI Agents

Learn the vocabulary for Agent configuration.

Lifecycle and Configuration

Design, test, distribute, and maintain an Agent.

DeepAgents

Understand multi-step workspaces and review controls.

Distribution Channels

Configure how users and systems reach an Agent.