3 Years of Graph Engineering with LangGraph
Cooper and Miles unpack LangChain's argument that “graph engineering” is not a new magic category, but a practical way to combine deterministic workflow control with agentic flexibility in LangGraph. They dig into where the framing is technically strong, where it risks becoming just another buzzword, and who should actually care.
Transcript
Cooper The funny part is, Miles, I wanted to roll my eyes at “graph engineering,” and then the article immediately agreed with me.
Miles Yeah, because the honest version is: the term is noisy, but the problem is not fake. Their central claim is that graphs are a reasonable way to put structure around non-deterministic software. Not because graphs are mystical. Because sometimes you actually know the valid paths before the model starts improvising.
Cooper Right.
Miles And that is very much our disease. Almost eight months of Exploring Next and somehow the boring edge between “let the model decide” and “absolutely do NOT let the model decide” is still where the action is.
Cooper But the hook here is not “graphs are back.” It is more like: prompt engineering, context engineering, harness engineering, loop engineering, now graph engineering. Builders keep inventing nouns for where they put the guardrails.
Miles Exactly. In LangGraph’s framing, nodes do work. A node can be deterministic code, a model call, a tool call, or a full agent with its own loop. Edges decide what happens next. And the strongest examples already have shape: support classification, coding agents inspecting repos, compliance approval before outside actions.
Cooper That’s where I buy it hard. The product story is not “graph-shaped agents are smarter.” It’s “you can encode the boring business reality directly.” And they back it with adoption, too: over sixty-five million downloads a month, maybe high sixties depending on the tracker.
Miles The technical part I liked most was their pushback against DAG thinking. Production agents usually need cycles. Retry a failed tool call. Ask for missing info. Revise after validation. Pause for a human and resume later. If your graph cannot loop, it is mostly a demo diagram.
Cooper There’s our content factory bingo card. Prompt, context, harness, loop, graph. I’m waiting for “vibes topology” by Friday.
Miles That matters because once planning, delegation, search, reading, and synthesis are open-ended, hardcoding the path can become premature structure. And I like that they explicitly say some systems belong in an agent harness like Deep Agents instead.
Cooper I agree, but the article is not saying graph all the things. My worry is downstream. Teams read “graph engineering” and start drawing boxes because boxes feel adult. If the graph does not expose decisions, approvals, retries, state, and receipts, it is just workflow cosplay.
Miles Exactly. And the “what’s actually new” section is subtle. Early nodes were deterministic code or one L L M call. Now a node can be a full agent run. Their docs-agent example has fixed operations, classifier and synthesis steps, then agents doing open-ended work in codebases.
Cooper That is the practical change. The unit inside the box got bigger. If your agent touches pull requests, internal docs, approval flows, customer threads, or billing-adjacent actions, you should care. Not because LangGraph is the only answer. Because the answer probably has explicit state and transitions somewhere.
Miles And dynamic transitions are the detail people should not skip. LangGraph has Send for cases where a node creates work at runtime. Research fans out across sources. The exact branches emerge from the task, but the contract around them is still visible.
Cooper So my honest take is: the argument holds if you read it as calibration. Graph engineering is not a new field. It is a name for a pressure every agent builder hits once demos meet policy, latency, cost, and review.
Miles Yeah… okay. I’m sold on that version. That admission about agent harnesses makes the graph case stronger, not weaker, because it marks the boundary.
Cooper No public bet from me here. The term might disappear, the pattern won’t. Ask me in six months whether we’re calling it graph engineering or, tragically, vibes topology.
Miles Please do not manifest that.
Cooper Too late, Cooper’s black-background landing page is mentally live. Alright, Miles. I’m closing three tabs and pretending that counts as graph garbage collection.