How to Choose Agent Frameworks for Reliable Production AI Workflows

What Happened

Recent releases put less emphasis on adding agent behaviors and more on making existing workflows recoverable, observable and safe to operate. LangGraph updated tool resumption, replay ordering and checkpoint packages; its Python SDK added streaming over SSE and WebSockets, with projections for messages and tool calls [2]. The OpenAI Agents SDK 0.24.0 expanded saved-run recovery state, improved session writes and tracing, and clarified how MCP tool annotations and tool errors should be handled [3]. Claude Code added a limit on how long unattended retries wait for usage to reset and expanded OpenTelemetry tool-decision events [1].

These releases do not establish a universal winner among LangChain, LangGraph, LlamaIndex, AutoGen, CrewAI, PydanticAI, DSPy, Semantic Kernel and the OpenAI Agents SDK. They do highlight the capabilities buyers should test: durable execution, controlled tool use, recovery, streaming and operational visibility.

Why It Matters to Businesses

An agent that calls external systems can leave work half-finished when a process stops, an approval pauses or a session write partially succeeds. A successful restart therefore requires more than replaying a conversation. It requires knowing which tools ran, which effects committed and what still needs human approval. The OpenAI Agents SDK notes that session-write batches can partially commit and that completed external effects must be reconciled rather than blindly replayed [3].

Kimbodo Engineering Perspective

Choose a framework around the workflow’s failure modes, not a demo’s number of agents. LangGraph is a candidate when explicit state transitions and checkpointing are central; its recent fixes to resume validation and replay ordering are directly relevant to that use [2]. The OpenAI Agents SDK is a candidate when its run, session and tool model fits the application, but its newer recovery snapshots cannot be read by older SDKs [3]. Claude Code is developer tooling rather than a substitute for an application runtime; its retry controls and traces matter when coding agents operate unattended [1].

Evaluate LangChain, LlamaIndex, AutoGen, CrewAI, PydanticAI, DSPy and Semantic Kernel against the same workload and acceptance tests. Avoid adopting multiple orchestration layers unless each has a defined responsibility.

How We Would Implement It

• Define a bounded workflow with explicit state, tool inputs and outputs, approval points, and terminal failure states. Keep authorization outside model-generated tool descriptions; MCP annotations are untrusted discovery hints, not permissions or schemas [3].
• Persist run state in a checkpoint store, such as Postgres where appropriate, and give each external action an idempotency key and an effect record. Test pause, resume and process failure; LangGraph’s Postgres and SQLite checkpoint packages gained conformance-suite testing [2].
• Put a policy-enforcing tool gateway between the agent and business systems. Require scoped credentials, validate arguments, and obtain approval before high-impact writes.
• Stream status to the user, but treat the durable run record—not the stream—as the recovery authority. Instrument model calls, tool decisions, approvals and retries with trace identifiers [1][2].
• Before upgrading an SDK, test snapshot compatibility and finish legacy runs with their original SDK and agent definitions where required [3].

Risks, Costs and Security

Long-running agents accumulate model, tool, storage and observability costs; set budgets and retry limits before enabling unattended execution. Treat tool metadata and retrieved content as untrusted, and verify sandbox boundaries: the OpenAI Agents SDK’s macOS local sandbox change does not provide complete host or network isolation [3]. Finally, test partial writes, stale concurrent updates and duplicate external effects. A checkpoint records progress; it does not make an external transaction atomic [3].

Sources

1. [1] v2.1.301
2. [2] langgraph-prebuilt==1.1.1
3. [3] v0.24.0