Three mindsets for building AI-agent systems — and they differ on one axis: how much structure you impose versus how much you let the model drive. A graph engineer hand-draws the workflow; a loop engineer designs the iteration and its stop; a harness engineer builds the model's environment and gets out of the way.
Graph— you draw the flow —Loop— you set the cycle —Harness— the model drives —
Not rival camps — a dial. The same engineer moves along this dial per task: impose more structure when you need guarantees, less when you want the model's full capability. Most real systems blend all three.
The axis The control ↔ autonomy spectrum
Each role sits at a different point between you controlling the flow and the model controlling it:
Golden rule: impose the least structure that still gives you the control and guarantees the task needs. More structure buys predictability; less unlocks the model's capability.
Most structure The graph engineer
🗺️ Graph engineer explicit nodes · edges · state
Builds The agent as an explicit graph: nodes (steps), edges (control flow, including branches and cycles), and a shared state object. You decide every route; the model fills in the steps you allow.
Best for Well-understood, structured, high-stakes or compliance-bound flows where you must guarantee the path — and where determinism and debuggability matter most.
Trade-off Rigid and more to build; struggles with the unexpected; gives the model the least autonomy. Tooling: LangGraph (see agent kits).
Some structure The loop engineer
🔁 Loop engineer trigger · cycle · stop
Builds The agentic loop: a trigger, what happens each cycle (reason → act → observe), and the stop criteria. The model iterates on its own until "done." Control lives in the loop's shape and its stopping conditions — not in the route.
Best for Open-ended tasks with a measurable stop ("make the tests pass", "until no new items") — where the number of steps can't be predicted.
Trade-off Less predictable than a graph; needs solid stop criteria, turn caps and guardrails to avoid runaway. Tooling: the agent loop & goal/loop primitives (see Getting started with loops).
Builds The environment around the model — context management (what it sees), tools, permissions, feedback and verification — then lets the model drive autonomously. The bet: a great harness + a capable model beats hand-coded control flow.
Best for Broad, open-ended, evolving work where you want the model's full capability — coding agents, research, general assistants. It's the frontier: context engineering and harness design.
Trade-off Least predictable and hardest to constrain; leans heavily on model quality and strong verification. Tooling: agent SDKs and Claude Code-style harnesses.
Why the harness bet keeps winning — the "bitter lesson"
Rich Sutton's bitter lesson: over time, general methods that scale with compute and data beat hand-crafted structure. Applied to agents: as models get more capable, elaborate hand-drawn control flow becomes a ceiling. A thin, powerful harness — great context, tools and verification — lets a strong model do more than any graph you could draw. So effort shifts from encoding logic toward shaping the model's environment.
Side by side
🗺️ Graph
🔁 Loop
🛠️ Harness
You control
The exact route
The cycle & stop
The environment
Model autonomy
Low
Medium
High
Determinism
High
Medium
Low
You build
Nodes, edges, state
Trigger, cycle, stop criteria
Context, tools, verification, permissions
Tooling
LangGraph
Agent loop · /goal · /loop
Agent SDKs · Claude Code harness
Best for
Known, high-stakes flows
Open-ended w/ clear stop
General, capable agents
Main risk
Rigidity
Runaway / loose stop
Unpredictability
In reality They compose — and the trend
These nest. A production system is often a harness that contains loops, some of which run small graphs for the parts that must be deterministic:
The direction of travel: as models get more capable, engineering effort shifts rightward — from drawing graphs, to shaping loops, to engineering harnesses and context. You hand more to the model and spend your effort on giving it a great environment and verifying its work.
How to choose today. Need a guaranteed path (money, compliance)? Put that part in a graph. Open-ended with a clear "done"? A loop. Broad and capability-bound by the model? Invest in the harness — and always pair autonomy with verification.
Cheat sheet
Question
Short answer
The one axis?
How much structure you impose vs how much the model drives.
Graph engineer builds…
Explicit nodes/edges/state — you control the route (LangGraph).
Loop engineer builds…
The trigger, cycle and stop — the model iterates until done.
Harness engineer builds…
The environment: context, tools, verification — the model drives.
Which is most deterministic?
Graph. Most autonomous? Harness.
How do they relate?
They compose (harness ⊃ loops ⊃ graphs); effort trends toward harness as models improve.
Knowledge check
Six questions on the three mindsets. Pick an answer for each, then check your score.