Back to glossary

AI/Agent Framework

YZ-Ajan Geliştirme Çatısı

D3

An AI/agent framework is a reusable software foundation that provides structures and components for building, running, or coordinating AI agents in an application.

Review status: 2026-12-04

Technical explanation

Frameworks can package choices about agent creation, models, tools, state or memory, sessions, evaluation, observability, and orchestration so a developer can assemble an agentic application.

Conceptual boundaries

A framework is software infrastructure, not the agent’s governance program. It is also distinct from an interoperability protocol such as MCP and from any assurance that an agent is autonomous, safe, compliant, or effective.

Provider-neutral example

Illustrative: a developer uses a framework to define a research agent, give it two read-only tools, attach session state, set an evaluation case, and route risky actions to human approval.

Limitations

Framework features and terminology differ by release and provider. Selecting one does not replace threat modeling, access control, evaluation, monitoring, or organizational accountability.

Related concepts

Atomic claims and evidence

  1. 1.1Johnson and Foote, Designing Reusable Classes
    Source
    Johnson and Foote, Designing Reusable Classes
    Source role
    Authoritative source
    Exact locator
    Introduction; ‘Toolkits and Frameworks’
    Supported claim
    Johnson and Foote describe a software framework as a set of classes embodying an abstract design for solutions to a family of related problems, supporting reuse at a larger granularity than individual classes.
    Last verification
    Review due
    Scope limitation
    This is a 1988 object-oriented software-framework definition. It grounds the reusable-software sense of this entry but does not define every modern AI or agent framework feature.
  2. 2.1Google Agent Development Kit documentation
    Source
    Google Agent Development Kit documentation
    Source role
    Authoritative source
    Exact locator
    Documentation navigation: Runtime, Components, Agent context, Evaluation, Safety and Security
    Supported claim
    Google ADK documents runtime, tools, skills, context, sessions, state, memory, evaluation, and safety/security, while Microsoft documents an agent framework platform for creating agents and incorporating agentic patterns into applications.
    Last verification
    Review due
    Scope limitation
    These are two independent providers’ documented product scopes. They support the AI/agent specialization of the entry, not a feature-completeness, governance, or performance claim for every framework.
    2.2Microsoft Learn: Semantic Kernel Agent Framework
    Source
    Microsoft Learn: Semantic Kernel Agent Framework
    Source role
    Authoritative source
    Exact locator
    ‘Semantic Kernel Agent Framework’
    Supported claim
    Google ADK documents runtime, tools, skills, context, sessions, state, memory, evaluation, and safety/security, while Microsoft documents an agent framework platform for creating agents and incorporating agentic patterns into applications.
    Last verification
    Review due
    Scope limitation
    These are two independent providers’ documented product scopes. They support the AI/agent specialization of the entry, not a feature-completeness, governance, or performance claim for every framework.
  3. 3.1Microsoft Learn: Semantic Kernel Agent Framework
    Source
    Microsoft Learn: Semantic Kernel Agent Framework
    Source role
    Authoritative source
    Exact locator
    ‘What is an AI agent?’ and ‘What problems do AI agents solve?’
    Supported claim
    Semantic Kernel documentation says agents can combine models, tools, human inputs, and other customizable components, and its framework supports modular components and process orchestration.
    Last verification
    Review due
    Scope limitation
    This describes software affordances, not a claim that automation should replace human judgment or that a framework supplies organizational governance.