What is an agent framework?
An agent framework is a software library that supplies ready-made building blocks for AI agents, such as the model loop, tool definitions, memory, multi-agent orchestration and tracing, so developers assemble agents instead of writing each part.
An agent can be built with nothing but a model API and a loop. A framework packages the parts most teams end up writing anyway: a tool-calling loop, typed tool definitions, conversation memory, ways to combine agents, approval steps, retries and tracing. The trade is speed and shared conventions against an extra layer between the developer and the model.
Anthropic's "Building effective agents" (December 2024) still suggests starting with model APIs directly, because frameworks can add abstraction that hides the underlying prompts and responses, and asks teams that do adopt one to understand the code underneath. OpenAI's practical guide prefers code-first frameworks over declarative graphs, and Google recommends its own ADK for agents on Google Cloud.
The landscape in late 2026
Lab SDKs
OpenAI Agents SDK, still pre-1.0; Claude Agent SDK, renamed from the Claude Code SDK in September 2025; Google ADK, whose 2.0 release reached general availability for Python in May 2026.
Graph runtimes
LangGraph, at 1.0 since October 2025 alongside LangChain, whose agent constructor now runs on it; Microsoft Agent Framework 1.0 (April 2026), the successor to Semantic Kernel and AutoGen, whose README now says it is no longer maintained.
Role-based multi-agent
CrewAI, with crews of role-playing agents and event-driven flows.
TypeScript toolkits
Vercel AI SDK 7 (June 2026) and Mastra 1.0 (January 2026), which builds on it.
Type-first Python
Pydantic AI v2 (June 2026), built around typed capabilities and a separate harness.
Framework, SDK, harness or runtime
The words overlap. A framework is a library of parts. A harness is the specific loop, tools, context handling and permissions that turn a model into a working agent, and several frameworks now ship one, such as the Claude Agent SDK, Pydantic AI's harness and LangChain's Deep Agents. A runtime is where agents execute in production, with durable sessions and sandboxes, such as Google's Agent Runtime or Claude Managed Agents.
All three labs now offer a raw API, an in-process SDK and a hosted runtime, which makes the real decision less about which framework and more about how much of the stack to own.
Where frameworks cause trouble
Hidden prompts
Defaults the team never wrote shape the model's behaviour, and debugging means reading framework source.
Version churn
Several frameworks shipped breaking major versions within a year. Tutorials written for older versions are often wrong.
Lock-in
Lab SDKs fit most smoothly with their own models and hosted state. Neutral frameworks avoid that but add their own abstractions.
Abandonment
AutoGen's README now says it is archived, with users pointed to Microsoft Agent Framework. A framework's maintenance status is part of the choice.
How Motiontic's engineers apply it
Motiontic is vendor-neutral. Motiontic's engineers pick per project, often starting close to the model API and adopting a framework where its parts, such as checkpointing or approval steps, would otherwise be rebuilt by hand. The deciding factors are the client's language and cloud, how much control the agent loop needs, and how well the framework exposes what is sent to the model.
Types of agent framework compared
| Lab SDK | Graph runtime | Role-based multi-agent | |
|---|---|---|---|
| Examples | OpenAI Agents SDK, Claude Agent SDK, Google ADK | LangGraph, Microsoft Agent Framework workflows | CrewAI |
| Model choice | Tuned for the lab's models, some support others | Model-neutral | Model-neutral |
| Control over steps | Model-led loop with hooks | Explicit nodes, edges and state | Roles and tasks, plus flows |
| Fits | Teams standardised on one lab | Long-running, auditable processes | Fast multi-agent prototypes |
| Main trade-off | Lock-in to one provider's stack | Verbosity and a learning curve | Less fine-grained control |
Common questions
Does a team need an agent framework to build an AI agent?
No. Anthropic recommends starting with model APIs directly, since many patterns take a few lines of code. A framework helps when a project needs durable state, approvals, multi-agent coordination or tracing that would otherwise be built by hand.
What are the main AI agent frameworks?
As of late 2026: LangChain and LangGraph, Microsoft Agent Framework, CrewAI, the OpenAI Agents SDK, the Claude Agent SDK, Google ADK, Pydantic AI, Mastra and the Vercel AI SDK, among others.
What is the difference between an agent framework and an agent harness?
A framework is a library of building blocks. A harness is the assembled loop, tools, context management and permissions that make one agent work. Many frameworks now ship a ready-made harness as one of their parts.
Related
LangChain vs LangGraph
The agent layer and the runtime underneath it, compared after the 1.0 releases.
Claude Agent SDK
Claude Code's harness as a library: the same loop, tools and context management, in a team's own process.
OpenAI Agents SDK
OpenAI's code-first agent framework: a small set of primitives that run in a team's own app.
Google ADK
Google's open-source agent framework, now with graph workflows, in five languages.
Microsoft Agent Framework
Microsoft's one forward path for agents, replacing Semantic Kernel and AutoGen.
AI agent development
Agents that do real work for users, with the evals, traces and guardrails to keep them in production.
Sources
- Anthropic, Building effective agents (2024-12-19)
- OpenAI, A practical guide to building agents (PDF, April 2025)
- LangChain blog, LangChain and LangGraph 1.0
- Microsoft, Microsoft Agent Framework Version 1.0 (2026-04-03)
- AutoGen README (archived notice)
- Claude Agent SDK docs, Overview
- OpenAI Agents SDK README
- Google ADK documentation
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.