Graph engineering

You draw the edges

Lesson 0006 asked who owns control flow: your path, or a loop. Graph engineering is the mix. You write the valid paths as nodes and edges so the model cannot skip classify-before-till — and a node may still be a whole agent when the work inside is open.

1. 1 · NodeDoes work. Code, one LLM call, or a full agent (0011).
2. 2 · EdgeWhat happens next. Fixed, or conditional on state.
3. 3 · StateThe ticket moving through. A graph is a state machine.

You already chose workflow vs agent. Tonight you may put both on one floor plan.

Architecture first. You imposed the paths. The model reasons only where you left a hole.

A loop is already a tiny graph

Khourshid via LangChain: a loop is a directed cyclic graph. LangChain’s simple agent loop is built on LangGraph. Loop engineering (0012) is not the opposite of graphs — it is the one-cycle case: act → verify → stop? → kick. Graph engineering is when you need more than that cycle: classify always before refund; inspect always before patch. Same essay · “loops are simple graphs”

Agent graphs are usually not DAGs: retry, ask a human, tool until enough. Cycles are the point.

# OPEN LOOP — model picks; may refund first
while not done:
    action = llm(messages, tools)

# GRAPH — you drew the edges
classify = code(ticket)              # deterministic node
if classify == "allergy":
    refund = code + harness stamp    # still 0007 on the till
elif classify == "recipe":
    sop = Agent(explore, brief)      # agent node (0011)
answer = llm("synthesize", sop)      # one call, no tools

# NOT A GRAPH
"Who else does walnut crumb?" → harness loop
# you cannot name the search path in advance

Kitchen only to make the floor plan concrete. The load-bearing picture is the state machine.

When not to draw edges

Open-ended research: plan, delegate, search, read, synthesize in orders you cannot pin. LangGraph’s team built early deep research as a workflow, then moved to an agent harness (Deep Agents) so planning and context stay in the loop. Forcing fake stages is how you get a graph that lies. Same essay · “When not to use graphs”

Skill: name the shape

Hope is an open loop that refunds first. Edges force classify. Mix puts a helper inside a node. Research should stay a harness.

Layers, not religions

Parked: durable execution, checkpoints, Send / map-reduce, LangSmith. A node that is “a whole coding agent” is newly practical — that is the 2026 twist, not the invention of graphs. Plugins still wait. Next: whether the shape moved a number you care about.

Check