Framework choice gets debated like a religious question and decided like one too, whatever the team already knows. That's a reasonable starting point, but it's worth being deliberate about what you're actually optimizing for before you commit a production build to one.
§ 01What framework choice actually changes
Not your agent's intelligence, that's the model, not the framework. Framework choice changes how much scaffolding you build yourself: state management, tool-calling conventions, checkpointing, observability hooks. A framework saves you from reinventing that scaffolding. It doesn't make a weak model smarter.
§ 02The main options, by what they optimize for
- LangGraph, graph-based orchestration, strong for complex multi-step and multi-agent workflows, large ecosystem, requires care around state-store security.
- Pydantic AI, type-safe, minimal-abstraction, built by the team behind Pydantic; a strong fit for teams that want structured outputs and fewer framework opinions imposed on them.
- Direct API orchestration, no framework, still a legitimate choice for a single, narrowly scoped agent; a framework's overhead isn't always worth paying for one workflow.
§ 03What to weigh beyond the benchmark
- How the framework handles state and memory persistence, and what security posture that requires from you.
- How much of your team already knows the framework's underlying language and conventions.
- Whether you need multi-agent orchestration now, or are over-building for a single-agent use case.
- Community and vendor support, a framework with a small, quiet community is a maintenance risk over an 18-month horizon, regardless of its technical merits today.
§ 04The mistake we see most often
Choosing a heavyweight multi-agent framework for a workflow that's genuinely one agent doing one job. The orchestration overhead adds failure surface, more moving parts, more state to secure, more places for a bug to hide, without adding capability the workflow actually needs. Match the framework's complexity to the workflow's complexity, not to what looks impressive in an architecture diagram.
§ 05Security posture differs by framework, and that's underweighted
The checkpoint and state-persistence layer is where recent production incidents have concentrated, not the model-facing layer most teams scrutinize carefully. Whatever framework you choose, budget real security review for its state store, not just its prompt-handling. We cover the specific failure patterns in our Lab.
§ 06The decision rule
Single agent, single workflow, small team → direct API orchestration or a lightweight, type-safe framework. Multiple agents, complex handoffs, need for durable long-running state → a graph-based orchestration framework, with a real security review of its state layer as a required, not optional, part of the build.
──
Not sure which framework fits your use case? A 30 minute discovery call will tell you, based on the actual workflow, not a benchmark leaderboard. See what a custom agent costs first if you want the numbers.