A feature checklist does not reveal whether a framework fits your workflow.
The solutionSelect orchestration and operational controls before selecting a brand.
Five frameworks worth evaluating in 2026
The right framework depends on what needs to be controlled: state, tools, retrieval, specialist delegation, or deployment. These are workload recommendations rather than benchmark rankings.
| Framework | Good fit | Key trade-off |
|---|---|---|
| LangGraph | Stateful workflow graphs, review steps, and resumable processes | State and graph design require engineering discipline |
| OpenAI Agents SDK | Code-first agents, function tools, handoffs, and tracing | Runtime guardrails are not blanket authorization for every hosted tool |
| LlamaIndex | Retrieval, knowledge-centric workflows, and document context | Retrieval permissions and source quality become critical |
| CrewAI | Role-based agent coordination and flow-style automation | Teams add complexity unless responsibilities are sharply bounded |
| Microsoft Agent Framework | Typed agents, workflows, middleware, and Microsoft integration | Check supported language and runtime features for your use case |
MCP is not a competing agent framework. The Model Context Protocol standardizes connections to tools and context. It can complement many of the runtimes above.
Which framework fits your workload?
Need to pause, resume, and approve?
Evaluate LangGraph or Microsoft Agent Framework. Test crash recovery, state transitions, and idempotent resumption.
Need a compact agent loop?
Consider OpenAI Agents SDK. Evaluate the exact tool types your application invokes and where pre-execution checks apply.
Need retrieval and grounding?
Consider LlamaIndex. Enforce access control on retrieved data and stop retrieved text from becoming trusted instructions.
Need role-based teams?
Consider CrewAI or an explicit workflow. Prefer one accountable executor over uncontrolled agent-to-agent tool access.
The production safety checklist
- Identity: whose authority is a tool using?
- Tool controls: can you restrict names, argument shapes, resources, and network destinations?
- Action boundary: can you inspect the final parameters immediately before execution?
- Approval: can a sensitive operation pause and resume without duplicate side effects?
- Observability: can you audit decisions without storing secret payloads?
- Failure handling: what occurs when a model call, tool, or policy service fails?
Agent Action Guard supplies optional Python framework integration helpers for LangChain, LlamaIndex, OpenAI Agents, CrewAI, and AutoGen, plus generic Python and JavaScript screening APIs. Confirm adapter coverage for your specific framework version and tool path. Guarding one function does not guarantee every external or hosted tool is screened.
2026 maintenance warning: AutoGen is archived
Microsoft’s AutoGen repository states that it is archived and no longer maintained. For a new Microsoft-stack project, investigate Microsoft Agent Framework and Microsoft’s AutoGen migration guide. Existing projects should assess security fixes and migration costs rather than assume the archived package will remain maintained.
Before picking any framework, build one end-to-end slice: a normal tool call, an unauthorized one, a human-approved one, a retry, and a visible failure. Choose the system that makes those paths clear.
Reviewed 3 November 2026. APIs and maintenance status can change. Follow the linked primary documentation for current versions.