Loop engineering is the name that stuck for something operators have wanted for years: software that keeps working on a goal after you close the laptop. Not a chatbot you prompt every morning. A system that builds, checks its own output against a real number, and tries again until it clears the bar, then does it again next week.
We have spent the past month rebuilding parts of our own operation, and several client systems, around this pattern. This is what we found, what it actually takes to run, and where we are pointing it first for operators of 10 to 50 person companies.
What a loop actually is
Strip away the buzzword and a loop is four parts: a task, a way to check the result, a target, and a schedule. The agent does the task. It checks the result against the target using an objective metric, not a vibe. If it falls short, it tries again with what it learned. If it clears the bar, it stops, or it waits for the next scheduled run and checks again.
Anthropic's own engineering team documented this exact shape as one of the core patterns for building agents: one process generates a piece of work, a second process evaluates it against criteria, and the pair loop until the output passes. They call it the evaluator-optimizer pattern, and it is worth reading if you want the underlying mechanics of why this works better than a single long prompt (Anthropic, Building Effective Agents). The business version is older than any of this. It is the build, measure, learn cycle that ran under lean manufacturing and later the lean startup movement. What changed is that an agent can now run every step of that cycle itself, including the measuring, instead of you doing the checking by hand.
Why a chat session is not a loop
Most operators' AI usage today is a chat window. You open it, you ask a question, you get an answer, you close it. There is no verify step and no stop condition. Nothing checks whether the output actually moved a number. That is fine for drafting an email. It is not how you run search rankings, ad spend, or a support inbox, because none of those improve from a single good answer. They improve from repeated small corrections measured against reality.
The technical piece that makes a real loop possible is tool access plus a place to delegate the checking step. Claude Code, for example, lets you define subagents. Each one runs a specific check (did the tests pass, does the page load, is the ranking up) in its own context and hands back a pass or fail (Anthropic, Claude Code subagents docs). Hooks fire an action automatically at a specific point in that cycle, such as re-running a check after every change, without a human triggering it (Anthropic, Claude Code hooks reference). That is the plumbing. The business logic, what to check and what target to set, is the operator's job, not the model's.
Three loops we are building for operators this quarter
We picked three because they map to where 10 to 50 person companies leak the most time and money, and because each one has a metric that already exists somewhere you can read it.
| Loop | What it checks | Cadence | Stop condition |
|---|---|---|---|
| SEO / content | Search ranking position for target keywords | Weekly or monthly | Page one, or your target position |
| Ad spend | Cost per lead against your breakeven number | Daily or weekly | Spend is at or below target CPL |
| Inbox / ops admin | Correct categorization and routing of incoming requests | Continuous, 24/7 | Accuracy holds above the threshold you set |
The SEO loop is the clearest example because the metric is unambiguous: where do you rank, and where do you want to rank. An agent can pull current position, compare it to competitors ranking above you, draft or revise the content, publish, and check again at the next scheduled run. This is not a five minute task. It is a loop that might run for months before it hits page one, taking one step at a time instead of one big rewrite. We run something close to this on our own blog already.
The ad spend loop works the same way against cost per lead or cost per booked call. Instead of a human checking the ad account every morning, the agent checks it, and only escalates to a human when the number moves outside the range you set. We build this as part of our automation work, wired directly into the ad account and CRM so the check is against real numbers, not a report someone has to pull manually.
The admin loop is the one most 10 to 50 person companies underrate. Every operator we talk to is losing real hours a week to manual routing, re-entering the same data across systems, or chasing a lead that sat too long. Our own research puts the typical admin leak at 10 or more hours a week per operator. That is a recurring, checkable task, and it is exactly what a loop is built for. You can see roughly where your own leak sits with our revenue leak heatmap. Speed to lead is a good example of an objective metric here: if a new inbound lead has not been contacted inside 5 minutes, the loop's check fails. It escalates or acts immediately, day or night, because the agent is not clocking out at 6pm.
What a loop needs to actually work
We keep seeing the same four requirements show up regardless of which part of the business the loop touches.
- An objective metric. Not "does this look better," but a number that comes from a system, not an opinion: a ranking position, a cost per lead, an accuracy percentage, a response time in minutes.
- Real tool access. The agent has to be able to act, not just describe what someone should do. That means it is connected to the CRM, the ad account, the inbox, or the CMS directly.
- A stop condition. Every loop needs an end state or it just burns time and tokens. "Rank on page one" is a stop condition. "Get better at SEO" is not.
- A schedule that matches the metric. Search rankings move over weeks, so a weekly check is enough. A lead response check needs to run continuously. Matching cadence to how fast the underlying number actually moves keeps the loop cheap and useful instead of noisy.
Running these loops is not free, but it is not expensive either. Claude access for a single seat runs from around 20 dollars a month on the standard plan, and heavier agent workloads scale on usage (Claude pricing). For a 10 to 50 person company, that is a rounding error next to the cost of a missed lead or a stalled ranking.
What to build this quarter
Start with one loop, not five. Pick the metric that already hurts the most, whether that is response time on inbound leads, ad spend that is not converging on a target, or a ranking that has not moved in months. Wire the check to a real number, give the agent the tool access it needs to act, and set a stop condition before you turn it on.
This is the same shape of work we do inside our AI Operations Agent and Second Brain builds: connect to the systems you already run, define the metric, and let the loop carry the repetitive verification work. That work used to sit on a person's desk every morning. The first version is usually live in days to weeks, not months, because it is wired into what you already have rather than replacing it. We have shipped similar systems inside client case studies. For teams that also handle a high volume of inbound calls, the same loop logic extends into voice agents that qualify and route callers against the same type of objective check.
If you are not sure which loop to build first, that is exactly what the AI Concierge assessment is for. It is a 999 euro engagement, and the full amount is credited to the build if you move forward, so the cost of finding out is not separate from the cost of doing something about it.
The point of loop engineering is not the term. It is that the checking step, the part that used to require a person opening a dashboard every morning and deciding what to do next, can now run itself against a number you trust. Set the target, wire the tools, and let it run.