SDK/library harnesses you build on — LangGraph + DeepAgents

The distinction between harnesses you use and harnesses you build on. LangGraph
is a production-grade runtime providing durable execution, streaming, and
checkpointing; DeepAgents is a batteries-included open-source harness for
long-running tasks, providing planning tools, sub-agents with isolated context, and
a virtual filesystem for prompts, skills, and memory. Students learn the
three-layer composition model — use the full harness out of the box, drop to a
lighter agent builder, or go all the way down to a custom graph — and that any
custom graph can itself be plugged in as a sub-agent. A guardrails note carries
forward to EB-8: such harnesses follow a "trust the LLM" model, so boundaries must
be enforced at the tool and sandbox level, not by expecting the model to
self-police. This topic is especially relevant for students building their own
product rather than adopting a tool.