Somewhere in a codebase you have never seen, an AI agent is having the game of its life. It has opened more pull requests this week than any human on the team. It resolved a customer case in forty seconds. It qualified a lead, drafted the follow-up, and logged the call before the coffee cooled. On every individual metric the dashboard offers, the agent is a star. And this, Salesforce argues in a late-June editorial, is exactly the problem.
The company’s framing is a stadium. During a World Cup, you thrill at the individual brilliance — the turn, the volley, the impossible save. But no matter how dazzling any single performance, the only thing that survives to the record book is whether the team won. Agents today, Salesforce writes, are “brilliant individual performers.” The uncomfortable follow-up is that individual brilliance and a winning result are not the same measurement, and for most of the last two years companies have been counting the wrong one.
What loop engineering actually means
The term the piece is selling is “loop engineering,” and to its credit the definition is refreshingly concrete. You hand an agent an objective and a way to measure progress. Then it does more than finish a task: it plans, checks its own work, learns from the result, and adjusts. That closed circuit — goal in, action, self-assessment, correction — is the loop. “That’s loop engineering at work,” the article says, and the phrasing is deliberate. The engineering is not in the agent’s cleverness. It is in the loop you build around it.
This is a subtle demotion of the thing everyone has been marveling at. A coding agent’s pull-request count, the piece points out, tells you the agent is busy. It does not tell you the software got better, the customer got served, or the quarter got closer to its number. The metric measures the footwork. It says nothing about the score.
The ultimate goal isn’t completing a task. It’s moving the business forward.— Salesforce, “Agents Run the Loop”
The evolution Salesforce sketches runs in three acts. First came agents that answer questions — the chatbot lineage, reactive and bounded. Then agents that act, reaching into systems to actually do the thing rather than describe it. And now, the piece claims, agents that pursue goals on their own: “running the loop, reading their own results, and adjusting.” Each act hands more autonomy to the machine, which is thrilling until you remember that autonomy without a scoreboard is just a very fast way to be confidently wrong.
You have already solved this problem
Here the article performs its central rhetorical trick, and it is a good one. Having spent several paragraphs convincing you that the hard part of agentic AI is measurement — knowing whether the outcome actually moved — it turns to the reader and says: “You have already solved this problem.”
The argument is that for twenty-seven years, Salesforce customers have been doing nothing but encoding their business goals into software. Every pipeline stage is a definition of progress. Every service level is a threshold of good. Every qualified-lead rule is a company writing down, in machine-readable form, what it means to be winning. All that CRM data, all that customer signal, all that automation and analytics accreted over nearly three decades — it was built to run a business, but it turns out to be the perfect infrastructure for an agent to optimize against. The business itself becomes the scoreboard.
It is, of course, a sales argument. The entire construction routes the reader toward the conclusion that the missing piece of their AI strategy is the platform they already pay for. But it is also, annoyingly, a reasonable point. The scarce resource in agentic systems is not intelligence, which is now rentable by the token. The scarce resource is a trustworthy definition of success, and a company that has spent decades arguing internally about what counts as a qualified lead has, without meaning to, built one.
The four systems, and the lifecycle
Autonomy that you cannot see is a liability, so Salesforce enumerates the scaffolding a loop requires. An agent needs four things: the controls to govern what it’s allowed to do, the context to ground it in your business, the traceability to see every step it took, and the analytics to know whether the outcome actually moved. Strip any one out and the loop breaks in a characteristic way — no controls and it’s dangerous, no context and it’s generic, no traceability and it’s unauditable, no analytics and it’s flying blind.
Wire all four together and you get what the company calls the Agent Development Lifecycle, or ADLC: a way to “see where it fell short, change the system that produced the miss, prove the change is better, and carry it forward.” It is, stripped of acronym, an admission that agents drift, that the first version is never the last, and that the interesting work begins after deployment rather than ending there. The loop is not just something the agent runs. It is something the humans run on the agent.
None of this is glamorous. It is testing, monitoring, calibration, and refinement — the unsexy machinery of keeping a probabilistic system honest over time. But it is the difference between an agent that impresses in a demo and one that survives contact with a real Tuesday in production.