Agent framework watchlist
A watchlist for teams moving from assistant usage into agent-backed product architecture. LangGraph represents explicit stateful orchestration; the AI SDK represents app-layer primitives. Together they frame the build-or-compose question: durable agent control versus smaller UI and model-call building blocks.
lane decision read
operator question
Does the product need durable agent control, or only app-layer AI primitives?
decision rule
Use this lane when assistant behavior must become repeatable product architecture with state, branches, and supervision.
avoid when
Avoid this lane when the job is a small AI feature, one-shot automation, or model call that does not need durable state.
compare by
Compare orchestration depth against product delivery speed and operational clarity.
tradeoff
Frameworks clarify complex agent flows, but they add surface area that simple AI features may not need.
ordered operator lane
Curated tools with metrics artifact signals
frontend-only composition
LangGraph
LGFAgent Frameworklangchain-ai/langgraph
signal
LangGraph matters when an assistant flow has become product architecture. The operator value is explicit control: states, branches, interrupts, retries, and human review points can be named and tested. The tradeoff is surface area. Reach for it when failures must be governed; avoid it when a direct SDK call or small workflow script would be easier to operate.
workflow fit
Best for product teams building stateful agent flows with explicit steps, branching, retries, and supervision.
watch out
Overkill for simple assistant features where a small SDK call or direct workflow script would be easier to operate.
Vercel AI SDK
VAISDKvercel/ai
signal
The AI SDK is most useful when the problem is shipping an AI product surface rather than researching agent control. It standardizes streaming, provider adapters, tool calls, and UI state so teams can move faster at the app layer. It does not replace decisions about memory, orchestration, permissions, or evaluation; compare it against agent frameworks when behavior needs durable control paths.
workflow fit
Best for TypeScript product teams shipping AI interfaces, streaming UX, tools, and structured output.
watch out
It supplies app primitives, not a complete agent operating model; architecture choices still belong to the product team.