My advice to design and eng leads chasing this, ignore the vocabulary churn, and ask your team one question, where in this system does a human or a verifier actually check the agent's work? If you can't answer that, no amount of renaming loops into graphs will save you.
"Are we still talking loops or did we shift to graphs yet?", Six Words from Peter Steinberger, and in 3 days the X timeline had produced courses, roadmaps, and thought-leader threads announcing that Microsoft, Stanford, and Anthropic had all simultaneously discovered graph engineering. 🤣
Hamel Husain followed with "Loop engineering is dead. Enter graph engineering." Both were half-jokes. Neither stopped the stampede.
I want to name what's actually happening here, because I've watched this exact cycle three times in the last eighteen months.
Prompt engineering -> context engineering -> harness engineering -> loop engineering -> graph engineering(5 days old) lol
Each rename gets treated like a paradigm shift. Almost none of them are.
The underlying problem hasn't moved. How do you keep an agent from lying to itself, drifting off task, or confidently doing the wrong thing across forty steps? Whether you draw that as a loop or a graph is an implementation detail. What matters is whether you've built explicit checkpoints, verifiers, and stop conditions, the graph framing is useful precisely because it forces you to make those things explicit instead of implicit in a single chat context. That's real progress. The renaming discourse around it is noise.
The moat in the agent era is harness quality, not model access. Model access commoditizes fast. A well-built harness compounds. Every serious team should be versioning their AGENTS.md the same way they version their code.
The coding agent leaderboard is not a model race. LangChain moved from 30th to 5th on Terminal-Bench 2.0 without touching the model. The gap was harness.
NexAU-AHE, automated harness optimization, lifts GPT-5.5 from 69.7% to 77.0% over 10 iterations with no human input. The harness is starting to optimize itself. The 'which model?' conversation ends when harnesses automate.
Andrej Karpathy just broke the entire premise of modern AI:
"Agents aren't magic. They're distillation at scale."
99.99% of your LLM's capacity is wasted on garbage data it never needed.
Small model + right tools + closed loop = terrifying capability.
In a 16-minute conversation, Karpathy reveals the full reasoning stack.
Worth more than any $500 AI course you've seen this year.
Police have finally started showing their ugly face to the protesters.
Just look at this young man.
He bravely confronted the police and challenged them to beat him, but he refused to bow down.
Gen Z revolution has flooded the streets🔥🔥
#CJPProtest