Peter Steinberger's July 18 post did not invent graph engineering, but it did expose the real fight: how much structure you need before an AI agent is safe to trust.
Peter Steinberger asked a small question on X and engineers treated it like a weather change. It landed. 'Are we still talking loops or did we shift to graphs yet?' was short enough to look like a joke, but it caught because teams are already making that decision in code. You either let one agent keep cycling through tools until it stops, or you draw the work as a graph of steps, gates, branches and approvals.
Start there. This is not a clean victory for graphs over loops. It is a vocabulary fight sitting on top of a practical engineering problem, and the practical problem is older than the phrase. If your agent can search, draft, test, retry, ask for approval and hand work to another agent, you already have a structure. The question is whether you designed it or merely allowed it to happen.
DeepLearning.AI's Agentic Knowledge Graph Construction course is part of why the subject has been circulating again, though the course itself is not brand new. DeepLearning.AI lists it as a free course taught by Andreas Kollegger of Neo4j, running 3 hours and 8 minutes, with lessons on using Google's Agent Development Kit to build a multi-agent system that turns structured and unstructured data into a knowledge graph. That is useful material. It is not proof that every agent pipeline now needs to become a graph.
That distinction matters. A knowledge graph represents relationships in data. A workflow graph represents relationships between tasks. The two can meet, but they are not the same thing, and mixing them together is how a sensible architecture discussion turns into jargon soup.
The loop still earns its place #
Anthropic's guidance is the right place to begin because it is deliberately plain. In its December 2024 essay Building Effective Agents, Anthropic said the most successful implementations it had seen used simple, composable patterns rather than complex frameworks. Its Claude documentation describes the agentic tool loop in direct terms: the model asks to use a tool, your application runs it, the result comes back, and the process repeats until the model stops.
That's a loop. It works.
A loop is cheap to understand and easy to change. You give the agent a goal, a tool set and a stop condition - plus enough logging to see what happened. For many product teams, that is all the architecture they need. Don't bother drawing a state machine when one bounded feedback cycle gets the job done and leaves a clear trace for review.
Anthropic's own multi-agent research system complicates the story in a useful way. In a 2025 engineering write-up, the company described an orchestrator-worker setup where a lead agent plans the search, spins up three to five subagents in parallel, and then combines their findings. Anthropic said parallelization cut research time by up to 90% on complex queries. It also said the system was token-hungry and needed explicit guardrails to stop agents from spiraling.
That's not trivia. It tells you the price of structure.
Graphs become useful when work branches #
Graph tools become more attractive when the work stops being one agent on one path. LangGraph's documentation points to persistence, checkpointing, human-in-the-loop review, time travel debugging and resumable execution. Google's ADK 2.0 documentation, released for Python in May 2026 and Go in June 2026, says the new workflow runtime evaluates agents, tools and functions as nodes in a graph. Microsoft has moved AutoGen into maintenance mode and now points new users toward Microsoft Agent Framework, which adds graph-based workflows for multi-agent orchestration.
Those are concrete product choices, not vibes. They all point in the same direction: once your system has branches, approvals, retries and shared state, you need a place to put that structure where engineers can inspect it.
Linear's Loops launch on July 20 is a cleaner example than most theory posts. Linear described Loops as recurring agent workflows for teams, available on Business and Enterprise plans, where a user writes the job in plain language and chooses whether it runs on a schedule or in response to an event. The agent can review issues, projects, documents, connected codebases and previous runs, then suggest or start work while keeping the run visible to the team.
Look closely at that design. Linear did not ship an open-ended robot wandering through a workspace. It shipped scheduled and event-driven work with visibility and control. That is the sane version of autonomy: a narrow task with an inspectable run and a human in the loop where it counts.
Frankly, the argument gets silly when people talk as if loops and graphs are rival religions. A graph can contain loops. A loop can sit inside one node of a graph. The useful question is smaller and sharper: does your workflow have enough branching state that a future reviewer needs to it, resume it, inspect it or approve a transition before it continues?
If yes, use a graph-style runtime such as LangGraph, ADK 2.0 or Microsoft Agent Framework. If no, keep the loop and spend your energy on good logs, solid stop conditions and a rollback plan. Simpler systems fail in ways you can understand. Overbuilt systems fail behind diagrams. The code review bottleneck makes this less academic than it sounds. AI tools can generate pull requests faster than human reviewers can clear them, and every team using them now has to decide where human judgment enters the pipeline. A graph gives reviewers explicit state transitions. A loop gives teams speed, but demands more discipline around observability and checkpoints.
Andrew Ng's course will not settle that choice. Steinberger's post will not settle it either. The value of this whole noisy week is that it names a decision builders were already making. You don't need to join a camp. You need to know which layer you are designing.
Also read: A Florida pastor nearly died following ChatGPT health advice and now his lawsuit could rewrite AI liability law | Small businesses are using AI to grow without hiring while big tech fires its way to the future | Genesis AI is in talks to raise $500 million at a $3 billion valuation just a year after emerging from stealth