·5 min read·loop-engineering

Loop engineering is a reliability problem, not an intelligence one

Agent loops are not failing because the models are weak. They are failing because nobody built the state, verifiers, gates, and receipts around them.

Contents

Linear called it "loop engineering" when it shipped Loops: the discipline of building agentic work that runs on its own, again and again, instead of once in a chat window. The name is right. The framing that usually follows it is not.

The instinct is to treat a struggling loop as a model problem: swap in a better model, write a longer prompt, add another example. Recent data says that is not where the failures are coming from.

Sinch's May 2026 global research, based on a survey of 2,527 senior decision makers across ten countries, found that 74 percent of enterprises have already rolled back or shut down an AI agent after deployment.

74%
of enterprises have rolled back or shut down an AI agent after deployment, per Sinch's 2026 AI Production Paradox research

The detail that matters most is buried one line down: the rollback rate climbs to 81 percent among organizations with mature governance frameworks. The best-governed teams roll back more often, not less, because they can actually see the failure. Everyone else is running the same broken loop and calling it stable.

#The gap is not intelligence, it is operating discipline

A loop that runs once in a demo and a loop you would trust with a client's operations are not the same artifact, even when they share the same model and the same prompt.

Gartner's June 2025 research draws the same line from a different angle: over 40 percent of agentic AI projects are expected to be canceled by the end of 2027, with escalating cost, unclear business value, and inadequate risk controls named as the drivers, not model capability.

40%+
of agentic AI projects are projected to be canceled by end of 2027, per Gartner's 2025 research on agentic AI

Put those two findings together and the pattern is consistent: the projects that survive are not the ones with the smartest agent. They are the ones where someone built the boring layer around it.

#What EP runs, and what we are not claiming

EP runs its own operation on a fleet of recurring agentic loops today: department seats that wake on a schedule, read their own state, do bounded work inside their lane, and report status back through a shared channel. That is a real, current fact about how we operate, not a demo.

What we are not claiming is a specific dollar or hour figure saved by running this way. That number does not exist yet in a measured, sourced form, so it stays out of this piece. A loop-engineering post that inflates its own numbers would be arguing against its own thesis on the way to making it.

The honest version of the claim is narrower and more useful: running our own operation on agentic loops is where we learned, first-hand, that the failure mode is almost never "the model got confused." It is state that lived only in a chat transcript, a claim of success nobody checked, a consequential action with no gate in front of it, or a run that left no trace of what it decided.

#The four things a loop needs before it earns trust

None of this requires a research lab. It requires treating the loop as a system, not a smarter conversation.

  • State outside the model. A cursor, a queue, or a ticket the loop reads before acting, so a restart or a different worker can pick the work back up without starting blind.
  • A verifier that is not the loop itself. A build, a test, a rendered-page check, or an independent read-back that can reject the loop's own claim that it worked.
  • Gates on the calls that matter. Routine, reversible steps run automatically behind a log. Anything expensive or hard to undo stops and routes to a person, with the uncertainty stated plainly, not smoothed over.
  • A receipt every time. What the loop did, what it verified, and what it decided, written somewhere a person can read later without replaying the whole session.

#The test before you trust a loop with real work

Before treating any agent loop as production-ready, ask what happens when it fails quietly, not when it succeeds loudly:

  • If this loop's state disappeared right now, could the next run recover?
  • If the loop claimed success, is there a check outside the loop that could have said no?
  • Is there a call in this loop expensive enough that a person should see it before it lands?
  • Six months from now, could someone read what this loop did last Tuesday without asking it?

A loop that cannot answer those four questions is not unreliable because the model is weak. It is unreliable because nobody engineered the system around it yet. That is the actual work, and it is available to any team willing to do it, with or without a frontier model.

See how EP builds loop engineering into every agentic system it runs.

#Sources

Keep reading