85 agentic tools audited for security incidents · Talk to the lab →
MotionticBook a call
Glossary

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Microsoft Agent Framework 1.0 (April 2026), the successor to Semantic Kernel and AutoGen, whose README now says it is no longer maintained.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

Types of agent framework compared
Lab SDKGraph runtimeRole-based multi-agent
ExamplesOpenAI Agents SDK, Claude Agent SDK, Google ADKLangGraph, Microsoft Agent Framework workflowsCrewAI
Model choiceTuned for the lab's models, some support othersModel-neutralModel-neutral
Control over stepsModel-led loop with hooksExplicit nodes, edges and stateRoles and tasks, plus flows
FitsTeams standardised on one labLong-running, auditable processesFast multi-agent prototypes
Main trade-offLock-in to one provider's stackVerbosity and a learning curveLess 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.

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.