Skip to content
Agent frameworks and protocols

How do you choose an AI agent framework?

First decide whether you need one: Anthropic recommends starting with the model APIs directly, because many patterns take only a few lines of code and frameworks can hide the prompts you need to debug. If you do need one, choose on how much control you need over each step, how state is stored and resumed, support for human approval and tracing, the languages your team uses, and whether it ties you to one model provider.

Published
22 October 2026
Sources checked
22 October 2026
From
Kramiva

01

Do you need a framework?

Frameworks save time on plumbing — tool calling, state, retries, tracing — but they add abstraction. Anthropic warns that frameworks can obscure the underlying prompts and responses and tempt teams into more complexity than the task needs. A single agent with a handful of tools is often easier to build and debug directly against the API.

02

What the main frameworks say they are

Each described from its own documentation, checked on the review date.

  • LangGraph: a low-level orchestration framework and runtime for long-running, stateful agents; it does not abstract prompts or architecture
  • CrewAI: Flows for structured, event-driven workflows and Crews for teams of autonomous agents, designed to be used together
  • Microsoft Agent Framework: agents and workflows for .NET and Python (Go in preview), the direct successor to Semantic Kernel and AutoGen
  • Google Agent Development Kit (ADK): model-agnostic, with sequential, loop, parallel and graph workflows, in Python, TypeScript, Go, Java and Kotlin
  • OpenAI Agents SDK: very few abstractions — agents, hand-offs, guardrails, sessions and tracing — with adapters for other model providers

03

Criteria that decide it

  • Control: can you see and change every prompt and every step?
  • State and durability: can a run pause for days and resume after a crash?
  • Human in the loop: can a person inspect and edit state before an action?
  • Observability: is tracing built in, and does it export to your tools?
  • Language and hosting: does it fit the stack your team already runs?
  • Model choice: can you switch provider without rewriting the system?

04

Trade-offs

This field changes quickly. Microsoft replaced AutoGen and Semantic Kernel with a new framework; Anthropic's own guide notes that tooling has moved on since it was written. Whatever you pick, keep your prompts, tools and evaluation set in your own code so a framework can be swapped without starting again.

Where Kramiva fits

Kramiva selects models by evaluation on the client's own task, builds integrations so the model can be swapped without rewriting the feature, and hands over the source code in the client's own repositories.

Sources

  1. Building effective agents — Anthropic, 19 December 2024
  2. LangGraph overview — LangChain
  3. CrewAI introduction — CrewAI
  4. Microsoft Agent Framework overview — Microsoft Learn
  5. Agent Development Kit (ADK) — Google
  6. OpenAI Agents SDK documentation — OpenAI

Accessed 22 October 2026. Kramiva is not affiliated with these publishers.

Related questions

Start

Tell us what you are building.

A short brief gets you a real reply from a founder within one business day — an honest read on fit, scope, and budget.

Start a projectExplore services