Graph engineering is the new term for designing the work around an AI, instead of pouring it all into one chat. Prompt engineering is the wording, context engineering is the inputs, and graph engineering is the shape. You draw jobs connected by arrows, some running in parallel, with a checking step and a human approval before anything expensive happens. The idea is sound, and the diagrams are useful.
Almost every version of the diagram has a skeptic node, a box whose only job is to attack what the researchers produced. None of them say which model sits in that box. The answer you reach by convention, the same model that wrote the work, is the one that breaks it.
What is graph engineering?
Graph engineering is designing an AI workflow as a set of connected jobs, instead of a single prompt. Each job has one owner and one output. Arrows show what has to finish first, and a shared record carries state between jobs. The words come from graph theory, yet the shape is how work already happens on a team.
The framing was popularised in a recent episode by Greg Isenberg on the Startup Ideas Podcast, and it lays out the starter shape most people draw first. A planner splits a question into lanes. Several researchers work in parallel, and a skeptic attacks the findings. A merge step turns the survivors into a recommendation, and a human approves before anything happens. The output is still a written report. What changed is that someone designed the work behind it.
When a graph earns its place
Save the structure for work with several steps, several sources, or paths that do not depend on each other. It also fits work that needs a grader, carries a real cost when wrong, or needs a human sign-off. Brainstorming names does not need a graph. Deciding whether to spend a year building something does.
This is a close cousin of the workflow patterns already documented in AI agent workflow patterns, where the parallel-then-merge shape shows up as fanout-and-synthesize. The name is new and the structure is old, and that is fine. A useful name for a thing people were already building has value.
The square the diagram leaves blank
Every graph-engineering diagram carries the same admission. The skeptic exists for one reason: the writer cannot be trusted to grade the writing.
A lot of AI research fails because the same model that writes the answer also grades the answer. That is like asking someone to write their own performance review and then being shocked when they describe themselves as a visionary. In a good graph, checking is its own job.
Greg Isenberg, Startup Ideas Podcast
That is correct, and it is where the argument stops. Checking gets its own box, the box gets a name, and nobody says what goes inside it.
Every worked example carries a checking step, across a support graph, a content graph and a coding graph. The point is always that the checker is not the writer. That is true, and it is not enough, because a different job can still run on the same model. A new job changes the org chart. A new model changes who is looking.
What the defaults actually do
This would be a small point if the tools made the other choice easy, and they do not. In every commonly named setup, the path of least configuration puts the same model in both boxes.
Level 1: manual lanes
You run each job as its own chat or doc, and you are the state. Whatever model you subscribe to answers the research prompt, and then it answers the skeptic prompt. The same weights run ten minutes apart, and nothing in the workflow tells you this is a problem.
Level 2: file trails
Each step writes its own file, so plan.md, customer.md and review.md leave a paper trail you can compare later. The trail is a real improvement. Still, the same assistant writes review.md and customer.md. The file names imply an independence the setup does not have.
Level 3: orchestrated
Frameworks give the skeptic a real node, with retries, state checkpoints and human-in-the-loop gates. There is no graph-level model setting for that node to inherit. That sounds like it should help, yet in practice it does the opposite. The model is whatever the developer typed into the node. The usual pattern is one model set up at the top of the file and reused in every node, the checker included.
So the diagram most people are drawing this week is a diamond. Its checker shares the writer's training data, stale facts and blind spots, and a measurable pull toward its own output. It has the look of review with none of the independence.
A new job title keeps the same reviewer
Splitting one model into a researcher step and a checker step changes the job title and keeps the reviewer. The same weights that produced the claim now judge it, and two separate effects push that judgment toward approval.
It prefers its own answers
In our own 80-prompt MT-Bench test, GPT acted as judge, and it picked its own output about 70% of the time, against a 33% baseline for impartial judging. Claude and Gemini came in near neutral at 32.5% and 31.25%. The published literature confirms the effect exists, yet it does not pin it to a clean number. So cite the direction of the effect, not the figure. More detail on why is in why AI can't check its own work.
It cannot report a tilt it does not notice
The sharper problem is a bias the model does not disclose. Work on value leakage (arXiv:2607.14345) found models quietly shaping answers toward their own preferences, while their stated reasoning claimed neutrality. A checker on the same weights inherits that tilt, and by design it cannot flag it. This argument for cross-vendor checking does not depend on quality rankings.
These two effects rank the vendors differently, so we name them separately. Self-preference and hidden value bias are different failures. A model that does well on one can do badly on the other. Merge them into one claim about which model is least biased, and the advice contradicts itself.
Notice the pattern. The skeptic node exists because one model grading its own work is unreliable, and then the default config hands that node back to the same model. TrueStandard does that step differently. Paste a draft, and four frontier models from four different labs check it in parallel, with every disagreement surfaced in sixty seconds.
Where a second vendor earns its cost
Using several vendors everywhere is the wrong lesson, because more agents bring more noise and more coordination cost. A graph that calls four labs at every node mostly pays to read four versions of the same paragraph. What matters is whether a node produces work or judges it.
| Node | Second vendor? | Why |
|---|---|---|
| Planner | No | Splitting a question into lanes is generation. One capable model is fine. |
| Researchers | Diversity, not judgment | Different vendors here widen coverage. Each brings different training data. Cheap and worth it. |
| Skeptic | Yes, strictly | The only node where the rule is absolute. It must never run on the vendor that produced the claim it is grading. |
| Merge | Prefer a neutral vendor | Merging decides what survives. Use a model that did not author most of the surviving claims. |
| Human gate | Not a model | The point of the gate is that a person decides. Adding a model here reintroduces the problem the gate exists to solve. |
Spend the second vendor at the skeptic node and you buy most of the available reliability. Spend it everywhere and you mostly buy latency, the wait for each answer.
The cheapest version of this is one extra call. Take the claim your research step is most confident about, hand it to a model from a different lab, and ask it to refute the claim instead of reviewing it. TrueStandard runs that pattern across four labs at once, the same move with four times the coverage.
Wiring the skeptic node
Four small changes, in rough order of how much they buy you.
1. Record who wrote each thing
Stamp the authoring vendor onto every research output, in front matter or a field. Without it you cannot route around the author later. People skip this step because it feels like bookkeeping, yet it makes the rest possible.
2. Exclude the author, in code
Make the routing rule real code instead of a note in a prompt. When the skeptic runs, keep only the models that did not produce the claim, and fail loudly if none are left. A rule you hope the prompt follows is not a control.
3. Ask it to refute, not to review
Review invites a summary, and refutation invites an attack. Tell the skeptic to default to rejecting when it is uncertain, and to end with one specific check a human could run this week. That last line is usually the most useful output of the whole graph.
4. Count votes against survival
A claim that came back unparsed or from a failed call has not survived. Treat a missing verdict as a rejection. The merge step then stops inheriting claims nobody checked.
One caveat worth keeping. Agreement across four vendors is the strongest signal this shape can produce, and it is still not proof. Models that share a blind spot agree confidently and are wrong together. That is why one AI checking another narrows the error bar without closing it. Cross-vendor checking leaves you less uncertain, and still short of certain.
That is the whole fix. The skeptic node was already in your diagram. It just needed a model that did not write the thing it is grading.
Frequently Asked Questions
Is graph engineering just a new name for agent orchestration?
Largely, yes. Directed graphs of jobs with dependencies are how workflow engines, business process modelling and DAG schedulers have worked for decades. Multi-agent frameworks have shipped this shape for years. The name is new and the structure is old, which is a fair objection. The name does add a reason to draw the workflow before you automate it. That catches the mistake of automating a process you have never run by hand.
Which model should run the skeptic node in an agent graph?
Any model from a vendor other than the one that produced the claim being checked. That rule matters more than which model you pick. Beyond that, prefer models that score near neutral when judging, instead of favouring their own output, because they make better graders.
Can I just run the same model twice with a critic prompt?
It helps with structure and obvious slips, and it is better than no check. It does not give you independence. Running the same weights again varies the wording rather than the underlying judgment, so a confident factual error tends to survive both passes.
What is the difference between a knowledge graph and an agent graph?
A knowledge graph maps how information connects, so a system can reason across relationships instead of retrieving the nearest-looking paragraph. An agent graph maps how work moves, so a task has steps, checks and handoffs. Graph engineering discussions usually mean the second, and mature systems end up using both.
How many models does a verification step need?
Two different vendors get you most of the benefit. The value comes from errors that do not line up, more than from headcount. Three or four raise confidence on claims that would be costly to get wrong. Past that, you mostly pay for agreement you already had.
Do LangGraph and similar frameworks default the skeptic to the same model?
Not by default, because there is no default to speak of. LangChain's docs are explicit that nodes are just functions, so nothing is inherited. Each node either hardcodes a model or gets one through runtime context, and that is why the outcome is so common. The framework makes no choice for you, and the convention is one model set up once and reused in every node. The checker node ends up on the same weights as the writer. Per-node model selection is fully supported, and nothing prompts you to use it.
Does cross-vendor checking make an agent graph much more expensive?
Less than most people assume, if you spend it in one place. Adding a second vendor at every node roughly multiplies cost. Adding one at the checking step alone adds a few calls and captures most of the reliability gain, because that is the only node where independence changes the answer.
Is graph engineering worth learning?
The vocabulary is worth an afternoon. Draw one workflow you already run as jobs and arrows, then run it by hand once, and that exercise pays. Build a large automated graph before you run the small manual one, and you produce mediocre work faster.
Keep reading
Multi-Agent vs Multi-Model AI in 2026
Multi-agent AI runs many roles of one model. Multi-model AI runs different models from different labs. Only multi-model can catch an error the first model is blind to.
Which AI Uses the Most Em Dashes? Ten Models Counted
Claude Sonnet 5 used the most em dashes of the ten models we tested: 9.55 per 1,000 words, more than twice the rate of GPT-5.6 and Gemini. We counted them, and 22 other AI writing tells, in 1,049 drafts.
Does ChatGPT Still Use Em Dashes? GPT-5.6 and GPT-6, Counted
We gave OpenAI's GPT-5.6 Terra and Luna 35 everyday writing jobs, three times each, and ran Luna at all five effort levels. Then we counted every em dash and 22 other AI writing tells. In October we added GPT-6 Luna and GPT-6.1 Sol.
Does Claude Use Em Dashes? Sonnet 5 and Haiku 4.5 Counted
We gave Claude Sonnet 5 and Haiku 4.5 the same 35 writing jobs as GPT and Gemini, three times each. Then we counted every em dash and 22 other AI writing tells.
Does Gemini Use Em Dashes? Fewer, but More AI Words
We gave Gemini 3.7 Flash and Gemini 3.8 Flash the same 35 writing jobs as Claude and GPT, three times each. Then we counted every em dash and 22 other AI writing tells. In October we added Gemini 4 Argon, and in November three more Gemini 4 models.
The Checker Shouldn't Be the Writer
TrueStandard is the skeptic node, run properly. Paste your draft, and four frontier models from four different labs check it in parallel and surface every disagreement in sixty seconds.
TrueStandard is for writers who publish AI-assisted drafts. Paid plans check every claim and citation across four AI models before you hit publish.
Free: one check a day, no signup. Pro is $20 a month for 250 drafts.
Start Verifying →