“Which agent framework should we use?” becomes easier to answer once you identify the task, execution environment and responsibilities you want the software to manage. A product name alone cannot answer those questions.
There are two kinds of framework in this library. A decision framework helps a process owner choose work and assign responsibility. A software framework helps a developer implement execution. The business pilot and responsibility guides cover the first. This guide covers development tools.
Identify what you are comparing
| Term | Meaning | What still needs a decision |
|---|---|---|
| Model | Generates responses and proposed tool calls. | Tools, permissions, execution and assessment. |
| SDK | Programming interfaces and supporting components. | Application behavior and hosting. |
| Framework | Structures agent or workflow implementation. | Task rules, integration and controls. |
| Runtime | Executes work and manages progress. | Data handling and external action recovery. |
| Hosted product | Supplies an application or managed service. | Configuration, access and operational ownership. |
| Protocol | Defines communication between systems. | Which capabilities to expose and who may use them. |
These categories can overlap. An SDK can include a runtime. A platform can host applications built with several frameworks. Check what a specific product provides rather than expecting the name to define its scope.
Different providers offer different approaches
OpenAI’s Agents SDK supplies an execution loop with tools, delegation, validation controls, sessions and tracing. OpenAI documentation.
Anthropic’s Claude Agent SDK embeds the execution capabilities behind Claude Code in applications you operate. It differs from a client library that simply calls the Claude API. Anthropic documentation.
Google ADK provides agent development and orchestration tools. Microsoft Agent Framework connects agents and functions through explicit workflows as well as supporting individual agents. Google documentation, Microsoft documentation.
GitHub’s cloud coding agent is a product for repository work. Comparing it directly with a general-purpose SDK mixes different responsibilities. GitHub documentation.
The five provider guides explain what you would need to assess for each approach. They are orientation guides, not performance rankings.
Where LangGraph fits
LangGraph is an orchestration runtime for stateful work, with persistence, durable execution and interruptions for a person’s decision. It can combine predictable steps with model-directed steps. LangChain supplies higher-level agent abstractions; the names refer to different parts of the stack. LangGraph overview.
For a support process, you could keep verification, approval and final recording as explicit workflow steps while letting an agent choose investigation tools between them. That is an illustrative design, not a tested recommendation for a particular organization.
Persistence creates another responsibility: deciding how external writes behave when execution resumes. A stored checkpoint does not automatically prevent a duplicate refund or message.
Compare with one representative task
Use the same task and acceptance criteria for candidate tools. Ask each implementation to retrieve a case, handle missing tracking and prepare a draft. Introduce an access denial and an interrupted write.
Record which behavior is built in and which requires your code. A feature advertised as approval support is useful only when you can tie approval to the precise action, pause work and resume correctly.
Evaluate the operational work too. Who hosts the runtime? Where is state stored? How are credentials rotated? Can your team understand a failed execution without reading undocumented internals?
Choose with explicit tradeoffs
Avoid treating one demonstration as a benchmark. Compare maintenance effort, integration fit, recovery behavior and costs alongside task quality. Keep model choice separate where the tooling permits it.
Start with a direct API call or fixed workflow if that satisfies the task. Add a framework when its execution and state features solve a requirement you actually have. Test the simplest suitable candidate before introducing more agents.
Selection checklist
- Identify the category of each candidate.
- Use the same task, data and criteria.
- Test interruptions and denied actions.
- Document application controls you must implement.
- Assess hosting, state, maintenance and provider dependence.
- Record your decision and revisit it after meaningful scope changes.
Tools, MCP and integration boundaries ↗
Understand what a tool exposes, how MCP connects it and where your application must enforce access.
Prepared with AI assistance and checked against the linked documentation. Examples and numerical limits are illustrative unless stated otherwise. These guides do not report independent product testing. Check current documentation before choosing a tool.