<- all tokdocs

You Have Been Building Graphs This Whole Time

Watch on TikTok

View on TikTok ->

Every LLM app you have shipped is already a graph. A node calls a model, a node calls a tool, an arrow moves state to whatever runs next. "Graph engineering" is not a new paradigm you missed. It is a name for the diagram you already drew without labeling it. The one real upgrade is what a node is now allowed to be.

Loop vs Graph, Side by Side

The video runs a single split-screen animation for its full 15 seconds. The left panel is a LOOP: an AGENT sits at the center of a ring, cycling through four fixed stops labeled DISCOVER, PLAN, EXECUTE, and VERIFY. A CONTEXT bar fills as the agent progresses, and its status ticks from WAITING to RUNNING to PASSED. The tagline underneath reads "Loop engineering: one job, repeated until done. The agent decides what's next."

The right panel is a GRAPH: a top-down flow of labeled boxes joined by explicit arrows. Work enters a ROUTER, branches to a RESEARCHER or an ANALYST, converges on a WRITER, then passes through a REVIEWER, a HUMAN CHECK, and finally OUTPUT (marked APPROVED). Its tagline reads "Graph engineering: specialized nodes, explicit routing. The topology decides where work goes."

Loop Graph
Controls Behavior Coordination
Question it answers How should this job repeat? Who runs next, and with what state?
Who decides next step The agent The topology (the arrows)
Shape One node, repeated Many specialized nodes, routed

The Only Thing That Actually Changed

The creator's framing is deliberately deflationary. Per the video, "Honestly? Less than the internet is telling you." The real shift is not the graph itself. It is what a single node is permitted to contain:

  • A node can be its own loop -- draft, check, improve, until it is good enough.
  • A node can be its own agent -- with its own private context and its own decisions.
  • A node can be a plain tool call -- deterministic, boring, reliable.
  • A node can be a human -- an approval gate before anything ships.

The HUMAN CHECK box in the right panel is that fourth case made literal. A pill at the bottom of the graph spells out the full menu: AGENTS . TOOLS . ROUTERS . HUMANS. The graph itself stays dumb on purpose. It is only the wrapper that decides who runs next and what state they receive.

Loops Live Inside Graphs

The payoff arrives as an animation. The WRITER node on the graph side lights up and reveals a nested cycle inside it, captioned DRAFT . CHECK . IMPROVE. On-screen text reads "A LOOP CAN LIVE INSIDE A GRAPH."

This resolves a false debate. Loops and graphs were never competing approaches. A loop controls how work repeats. A graph controls how work moves between roles. One nests inside the other. The WRITER keeps looping on its own draft until quality clears, then the graph hands the state to the REVIEWER. The loop owns the local behavior. The graph owns the handoff.

The Practical Takeaway

The closing frame states the rule plainly: "Loops control behavior. Graphs control coordination." The advice for anyone already chaining prompts, tools, and other prompts is not to rebuild anything. It is to make the implicit structure visible:

  1. Draw the roles (what each step actually is: agent, tool, loop, or human).
  2. Draw the arrows (who hands off to whom).
  3. Name the state (what data moves along each arrow).

Do that and, in the creator's words, "the 'advanced architecture' everyone's posting about is just... your diagram." The footer caption puts it as an operating principle: "if you can draw roles and arrows, engineer the graph."

Key Takeaways

  • Any LLM app that chains models, tools, and other prompts is already a graph. Naming it does not change what you built.
  • The genuine upgrade in graph thinking is node flexibility: a node can be a loop, an agent, a tool call, or a human approval gate.
  • Loops and graphs are not rivals. A loop governs how one step repeats; a graph governs how work routes between steps. Loops nest inside graphs.
  • Loops control behavior. Graphs control coordination. Keep the two concerns separate in your head.
  • To adopt "graph engineering," draw the roles, draw the arrows, and name the state that flows between them.

Resources

  • No external resources referenced in this video.

Published July 26, 2026. Writeup generated from a favorited TikTok.