Part 10 of 12 · 1 min read
Quick notes
- A framework is a bundle of paradigms plus plumbing: a loop, a graph with checkpoints, and handoffs or agents-as-tools, plus tool schemas, tracing, guardrails and sessions.
- Compare on stable capabilities, not on names, because frameworks change fast: persistence, deployment, guardrails and observability. Many teams build directly on provider APIs.
- LangGraph checkpoints per-thread state, the Claude Agent SDK stops on no tool call or at a turn or budget limit and exposes hooks, and the OpenAI Agents SDK offers handoffs, agents-as-tools and guardrails (as documented on 31 October 2026).
- A checkpointed graph finishes more runs with fewer calls when steps can crash, but checkpointing alone still repeats side effects; only an idempotency key removes the duplicates.
- Handoff versus agent-as-tool is a context-cost question: handoff was cheaper with one or two specialists, agent-as-tool from three up. Nothing here says which framework is faster or more accurate.