Every couple of months, if you follow AI as closely as I do, a new buzzword arrives. First it was prompt engineering. Then context engineering. Then harness engineering.
The two latest are loop engineering and graph engineering.
So what exactly are loops and graphs?
Last week a CMO at a large law firm asked me to explain what a loop is. She apologized for asking. I told her to stop apologizing, because these terms get thrown around in meetings and on LinkedIn without anyone defining them.
I promise you can follow this. If you can read a subway map, you already understand the hard part.
The loop: doing something over and over
A loop is doing the same thing repeatedly until the work runs out.
Picture someone sealing 100 boxes on an assembly line. Seal a box, then move on to the next one. Keep going until the boxes are gone.
You already run loops in your marketing department every day. Think about reviewing every attorney bio on the website, one profile at a time. Or proofreading a fifteen-page client alert, page by page.
Each one is a repeated action applied to a pile of tasks, until the pile is empty.
When AI runs a simple loop, it follows a fixed recipe. It completes the same steps in the same order, every single time. Feed it 50 attorney bios and it will process bio number 50 exactly the way it processed bio number one.
Computers happen to be spectacular at this. A person sealing boxes gets tired, bored, and sloppy somewhere around box 60. A machine seals box 10,000 with the same precision it brought to box one, at three in the morning, without complaint. Whenever you hear the word "automation," this kind of loop is what the vendor means.
Simple loops carry one big limitation, though. They repeat without improving. Box 100 gets sealed with exactly the same skill as box one.
Feedback: the step that makes a loop learn
Add one step where the system checks its own work against what it was trying to produce. That is enough to make the loop learn.
Think about how a kid learns free throws. She shoots the ball and watches where it lands. When the shot falls short, she bends her knees more on the next attempt. When it clangs off the back rim, she softens her release. Shot after shot she adjusts, and by August she is draining ten in a row.
She ran a loop. Because the loop included a feedback step, watching the ball and adjusting her form, each trip through the cycle made her slightly better than the trip before.
Your thermostat runs the same kind of loop all day. It measures the room, compares the reading to your setting, nudges the heat up or down, waits a while, and measures again. Each individual step is dumb. Stacked into a cycle that repeats around the clock, those dumb steps keep your house at exactly 72 degrees.
You run feedback loops at work too. You send a newsletter, check the open rate, tweak the subject line, and send again next month, even if you never used the term for it. Anyone who has retired a tired headline formula because the numbers kept sagging has run a feedback loop.
The formula is always the same four beats: try, check, adjust, try again.
This is exactly how AI learned to write
The AI tools you use every day were built almost entirely out of one giant feedback loop.
A simplified version of how it works: the system reads a sentence from the internet with the last word hidden, then guesses the missing word. After guessing, it peeks at the real answer. A wrong guess triggers a tiny internal adjustment, so the next guess comes out a little closer. Then it moves on to the next sentence.
That single cycle ran billions upon billions of times, across a huge portion of the written internet. All of the grammar, the facts, the tone, and the legal vocabulary these tools display emerged from that one repeated exercise of guessing and correcting.
A second loop came afterward. Human reviewers read pairs of AI answers and picked the one they preferred. The system then adjusted itself toward the winning style. Thousands of rounds of that loop taught the tools to answer in the helpful, conversational way you experience today.
The kid at the free throw line and the chatbot on your laptop got good the same way. Both repeated a task and corrected course a little before the next attempt. Repetition plus feedback, at unimaginable scale, produced the tool that now drafts your client alerts.
Loops inside your workflows
The same feedback idea appears inside the AI workflows you might buy or build for your team.
Say an AI drafts a client alert about a new regulation. A basic loop hands you whatever comes out on the first pass, typos and awkward phrasing included. A feedback loop runs the draft through a checklist first, confirming that the citations exist and that the tone matches the firm's voice. A draft that fails a check goes back for another revision pass. The cycle continues until the draft clears the bar or hits a retry limit.
One more layer sits on top. When your team corrects the final output, those corrections can be written back into the workflow's instructions. Next month the loop starts from a smarter place. Over a year of this, a well-tended workflow begins to sound less like a robot and more like your firm.
A workflow with feedback built in compounds in value, the same way the free throw shooter compounds skill.
The graph: a map of connected things
Now for the second idea. A graph is a map of things connected to each other.
Think of a subway map. It shows stations with lines running between them. From any station you can travel in several directions.
Your LinkedIn network is a graph. Each person is a point, each relationship is a connection, and there are many possible paths from you to anyone else. A family tree works the same way, and so does an airline route map.
The defining feature of a graph is that every point connects to others, which creates choices about where work can flow.
Why one person juggling everything breaks down
To understand why graphs are such a big deal in AI, think about your own workday.
Say a partner announcement drops Thursday. The press release needs drafting, the LinkedIn posts need writing, the bio page needs updating, and the client email needs a final proof. One person handling all four has to juggle, and juggling has a cost. Each switch between tasks burns focus, and quality slips a little with every jump. Anyone who has proofread an email while half-listening on a Zoom call knows the feeling.
A good marketing director solves this with her org chart. She hands the press release to her PR manager, the social posts to her content specialist, the bio update to the web coordinator, and keeps the client email for herself. Four tasks now run at the same time, each getting someone's full attention. When the pieces come back, she reviews the whole package and ships it.
Look closely at what she just used. An org chart is a graph. She sits at one point on the map, her team members sit at connected points, and work travels along those connections in both directions.
Agents delegating to agents
AI runs into the same juggling problem she did. A single AI grinding through a long, complicated job in one sitting behaves like one employee doing everything alone. The longer the to-do list, the more likely it drops a detail somewhere in the middle.
A graph solves this through delegation. One AI sits at the center as the project manager. It breaks the job into pieces and hands each piece to a helper, which the industry calls a subagent. One subagent might research the partner's background while a second drafts the press release, with a third checking facts against source documents and a fourth formatting the bio for the website. Each helper gives its single narrow task full attention. Several of them can run at the same time.
When the helpers finish, their work flows back along the connections to the project manager. The manager assembles the pieces, runs the whole package through a final quality loop, and either drops it in your approval queue or flags a problem for a human.
That structure is what vendors mean by "multi-agent systems" or "subagents." The picture in their heads is an org chart where every box is an AI.
The advantage comes down to focus. A subagent with one narrow job keeps every relevant detail in view from start to finish, the same way your delegating director trusts each specialist to give one assignment undivided attention.
Putting it all together: what an "agent workflow" really is
An agent workflow is a graph with loops inside it.
The graph is the map: an org chart with forks, where work gets routed and delegated based on the situation. At various stops along the map, an AI cycles through drafting and checking until the task clears a quality bar. Those cycles are the loops.
The starting point stays the same every morning, but the path changes depending on what happens along the way. It works like a choose-your-own-adventure book, where the story branches based on decisions made at each fork.
An assembly line only ever repeats one fixed path. A graph offers many paths and chooses among them based on what it finds en route.
Why you should care
Routing is what makes agents powerful, and it is also where they fail. Every delegation is a handoff, every handoff is a place where instructions can get garbled, and nobody catches it until the output sits on a partner's desk.
Every graph workflow needs human checkpoints for exactly this reason. The AI does the execution; you decide which goals to pursue and sign off on the output.
Understanding these two ideas also lets you decode sales language. "Automation" usually describes a simple loop. "Built-in quality assurance" usually describes a feedback loop. "Agentic" or "multi-agent" describes a graph, which means more places for work to go wrong, so ask more pointed questions about the handoffs.
So the next time someone drops "agentic" into a pitch meeting, ask one simple question: where are the forks in the road, and who approves the turns?
If they stumble, you just learned they are selling you a conveyor belt with a fancy name.