Explainer

Loops vs graphs

What
Agent architecture: when a loop should become a graph, and when it shouldn’t. A test you can run on any agent you build.
Prompted by
Peter Steinberger’s July 2026 “loops or graphs?” question. Adapted from my own prototype on agent architecture; public-safe, no proprietary detail.
Back to
Explain

In July 2026 Peter Steinberger poked the field with one line: “are we still talking loops, or did we shift to graphs yet?” It read as a verdict. It wasn’t. It was a question, and the answer is a test you can run on any agent you build.

The one idea

A loop is one agent iterating on its own judgment until it’s done. A graph cuts that work into fixed nodes joined by gated edges. Escalate only to buy a gate, determinism, or a real handoff.

1The two shapes

These are not “old vs new.” A loop is a cycle; a graph is a fixed path with checkpoints. The defining difference isn’t how many agents there are. It’s who decides what happens next: the model’s own judgment (loop), or a structure you wired in advance (graph).

Loop — one agent, iterating
StepObserveRead the goal and current state.
↓
StepActPick one tool, run it, see the result.
↓
The cruxSelf-checkThe same agent judges “done?” — actor and judge are one.
↺ not done → back to Observe
Graph — fixed nodes, gated edges
Node ADo stage oneTyped state passes along the edge.
↓
The edge that earns itGateA hard check or human approval the flow cannot pass without.
↓
Node BDo stage twoOnly reachable once the gate cleared.

The crux node is the loop’s problem: it grades its own work. The gate is the only thing a graph gives you that a loop can’t: an edge the flow is forced through. Everything else a graph “adds”, more steps and more prompts, a loop already has.

2When a graph earns its place

There are exactly three honest reasons to break a loop into a graph. If none apply, you’re adding latency and rigidity for tidiness.

Gate
A step must be blocked until a hard condition passes or a human approves. A self-judging loop can’t enforce a stop it isn’t allowed to overrule.
Determinism
You need a boundary to behave the same way every time — a fixed route, a checkpoint you can resume from — not “whatever the model decided this run.”
Real handoff
Stages genuinely differ — different tools or models per node (cheap researcher → strong writer → strict reviewer), not one agent play-acting all three.
Not a reason
“The prompt got long.” “Four steps feels like a lot.” “A diagram looks more professional.” Step count and neatness aren’t reasons.
The collapse test

Try to fold every node back into one agent’s loop. If you can and lose nothing — no gate goes unenforced, no determinism disappears, no real tool/model split collapses — it was never a graph. It was a loop wearing a diagram, and you paid latency for the costume.

Rule of thumb: if five nodes collapse back into one loop and nothing is lost, you should collapse them. Master the loop first. Reach for a graph only when the work forces your hand.

3Check yourself

Six questions. The answer appears when you pick one.

Explain ·andrewshiau.com · Press ⌥G for the grid · Set on a 12-column grid