检索到的文档、内存、工具定义、会话历史。

retrieved documents, memory, tool definitions, session history.

• 𝐇𝐚𝐫𝐧𝐞𝐬𝐬 𝐄𝐧𝐠𝐢𝐧𝐞𝐞𝐫𝐢𝐧𝐠 (𝐅𝐞𝐛 2026): The constraints and gates that keep an agent in bounds. A good prompt won't stop an agent from rewriting your entire codebase if nothing architecturally prevents it.

• 𝐋𝐨𝐨𝐩 𝐄𝐧𝐠𝐢𝐧𝐞𝐞𝐫𝐢𝐧𝐠 (𝐉𝐮𝐧𝐞 2026): Designing the plan-execute-verify cycle an agent repeats, and deciding what triggers each step and what counts as done.

• 𝐆𝐫𝐚𝐩𝐡 𝐄𝐧𝐠𝐢𝐧𝐞𝐞𝐫𝐢𝐧𝐠 (𝐉𝐮𝐥𝐲 2026): Wiring multiple agent loops together through nodes, edges, and shared state, instead of relying on one agent to handle everything sequentially.

Lay them out like this and the picture gets clearer. Prompt, context, and harness engineering are genuinely different layers of the same single-agent problem. Loop and graph engineering sit much closer together, a graph is just what happens when one loop isn't enough and you need several coordinating.

And here's the part worth sitting with: harness to loop to graph is three renamed "eras" in about five months. That's a pace no other engineering discipline moves at, which says something about how much of this is new architecture versus new vocabulary.

We found 6 frameworks that were already doing "graph engineering," nodes, edges, shared state and all, before the term existed. Full breakdown on the 𝐛𝐥𝐨𝐠, link in the 𝐜𝐨𝐦𝐦𝐞𝐧𝐭𝐬.

#AIengineering #AgenticAI #LLMEngineering #AIagents #MultiAgentSystems #LangGraph