A single good answer is not a finished job. Real work is a loop. You do a step, look at the result, fix what is wrong, do the next step, and stop when the thing is actually done. People do this without thinking about it. When we ask AI to do the same work, someone has to design that loop on purpose. That design is called loop engineering, and it is quietly the difference between AI that demos well and AI you can rely on.
From asking a question to finishing a job
It helps to see where this fits. AI has gone through three phases, each one solving a problem the last one exposed.
- Prompt engineering is asking the right question. You phrase the instruction well and the model gives a better answer. This is where most people started.
- Context engineering is handing over the right information. You give the model the document, the client record, the policy, so its answer is grounded in your reality instead of a guess.
- Loop engineering is designing how the model actually does a multi-step job. Not one answer, but a cycle: take an action, check the result, decide the next action, and know when to stop. This is the phase that turns a clever answer into completed work.
A useful way to picture it: prompt engineering is asking a good question, context engineering is giving someone the right files, and loop engineering is how a capable employee actually gets through a task, doing a step, sanity-checking it, correcting course, and recognizing when the job is finished. The first two get you a smart response. Only the loop gets you the work.
What the loop actually is
When AI does a real task, it does not answer once and walk away. It runs in a loop. Say the job is “open a new client matter.” The loop looks something like this:
- Act. Read the intake email, pull the client details, check them against existing records, create the matter, send the confirmation.
- Observe. Look at what happened after each step. Did the record save? Was there a conflict? Did the email go out?
- Decide. Based on what it just saw, choose the next move, or retry, or flag a human.
- Stop. Recognize the job is genuinely complete, and end, rather than spinning or stopping halfway.
Loop engineering is the work of designing that cycle: what the AI is allowed to do, how it checks its own work, what happens when a step fails, and, crucially, when it should stop or hand off. Get the loop right and the task finishes correctly. Get it wrong and you get confident nonsense, quietly.
Why loops are hard to get right
The reason loop engineering is its own discipline is that loops fail in ways a single answer never does. A few of the traps:
- Errors compound. This is the big one. A step that is right 95% of the time sounds fine, until you chain ten of them together and the odds of getting all ten right drop to roughly six in ten. One wrong turn early poisons everything after it, and the loop keeps going as if nothing happened.
- Loops can run away. Without limits, an AI can retry the same broken step over and over, wander down a wrong path, or spin in circles. With tools priced per use, a runaway loop is not just slow, it is an unpredictable bill.
- The same input can take a different path. Left to improvise, the model may do the job one way today and another way tomorrow. That is hard to trust and harder to audit when something goes wrong.
- Knowing when it is done is its own problem. A loop that stops too early leaves work half-finished. One that never decides it is done keeps burning time and money. Both look like success from the outside.
- Something has to catch failures. A portal changes, a document is malformed, a field is blank. Does the loop notice and recover, or does it sail past and produce a clean-looking wrong result?
Why this matters for the business, not just the engineers
It is tempting to file this under “technical detail.” It is not. The loop is exactly where the value and the risk live. A demo has to work once, for the person who built it, on the example they chose. A loop you put into production has to work a thousand times, on the messy inputs nobody chose, without a human watching each run. That gap is the whole game.
And in back-office work, the failure is not a chatbot giving a weird answer you can shrug off. A loop that quietly gets a step wrong sends the invoice with the wrong number, misses the conflict, drops the client email, or lets the deadline slip. No one reviews it, because the point was to take it off someone's plate. A well-engineered loop finishes the work correctly every time, at a cost you can predict, without anyone babysitting it. A poorly engineered one erodes trust until the firm concludes “we tried AI and it did not work,” when what actually failed was the loop.
The tight loop: how Caddi handles it
The fix is not a smarter model left to improvise. It is a tighter loop: bounded, predictable, and owned. That is the whole idea behind Caddi.
Caddi is an agent that learns how you work, the way you would onboard a new hire. It watches a screenshare, reads your written instructions, and looks at the tools and data you already use. From that it builds a fixed, deterministic loop, a known set of steps that run the same way every time, rather than an agent reinventing the task on every run. Because the steps are locked in and run over APIs instead of by clicking around a screen, the loop does not wander, does not spin up a surprise bill, and does not shatter the moment a vendor moves a button. Checks and failure handling are part of the loop, so a bad input gets flagged instead of quietly turning into a wrong result.
Two more things make it a real option rather than an experiment. The cost is flat and predictable, not metered by how chatty the model felt that day, so finance can plan around it. And Caddi maintains the loop for you, so when a system changes, keeping it running is a vendor's job, not a favor someone squeezes in between real work. You get the finished work without owning the loop. For the longer version of why in-house builds break on exactly these points, see why lawyer-built AI dies before it reaches the firm.
Loop engineering is going to be a phrase you hear a lot. Underneath it is a simple question you can ask any AI tool: when it does your work, is the loop tight enough to trust unwatched? With Caddi, that is the part we build for you.
Skip the fragile loop
See a tight loop built for real work
Tell us where to reach you and the calendar opens right here. In 30 minutes we'll show you how Caddi learns your workflow from a screenshare, your notes, and the tools and data you already use, then runs it as a bounded, deterministic loop that finishes the same way every time, at a flat cost, maintained for you.
Frequently asked questions
What is loop engineering?
Loop engineering is the discipline of designing how an AI system does a multi-step job, rather than how it answers a single question. Real work is a loop: take an action, look at the result, decide the next action, and stop when the task is actually finished. Loop engineering is defining what the AI is allowed to do, how it checks its own work, what happens when a step fails, and when it should stop or hand off to a human. It is the phase that turns a clever answer into completed work.
How is loop engineering different from prompt engineering and context engineering?
They solve three different problems. Prompt engineering is asking the right question so the model gives a better answer. Context engineering is handing the model the right information (documents, records, policies) so its answer is grounded in your reality. Loop engineering is designing the multi-step cycle the model runs to actually finish a job. The first two get you a smart response; only the loop gets you completed work.
Why do AI agent loops fail?
Mostly because errors compound and loops are hard to bound. A step that is right 95% of the time seems fine, but chain ten of those together and the odds of getting all ten right fall to roughly six in ten, and the loop keeps going as if nothing went wrong. Loops can also run away (retrying a broken step, wandering, spinning), take a different path on the same input from one run to the next, stop too early or too late, and sail past failures like a changed portal or a malformed document, producing a clean-looking but wrong result.
How does Caddi make the loop reliable?
Caddi replaces an open-ended, improvising agent with a tight, bounded loop. Caddi is an agent that learns how you work, the way you would onboard a new hire: it watches a screenshare, reads your written instructions, and looks at the tools and data you already use. From that it builds a fixed, deterministic set of steps that run the same way every time, over APIs rather than by clicking around a screen. It does not wander, does not run up an unpredictable bill, and does not break when a vendor moves a button. Checks and failure handling are built into the loop, the cost is flat and predictable, and Caddi maintains the loop for you, so it keeps finishing the work correctly without a human watching every run.