AI Agent Loops Are Here. Enterprises Need Budgets and Bosses for Them

For most of the past two years, using an AI agent meant asking it to do something and waiting for an answer. This week brought fresh signs that the next phase looks different: agents that keep running in the background, wake themselves up on a schedule and keep working until a goal is met. Developers have started calling these "loops." For enterprise leaders, loops promise steady, unattended progress on work like code maintenance and support triage. They also raise questions most governance programs were not built to answer, starting with who is watching and how much it costs.

What are AI agent loops?

TechCrunch AI editor Russell Brandom described the trend on June 22 as a swarm of agents working continuously in the background. The idea borrows from a basic programming concept, the recursive function that keeps calling itself until a stop condition is met. The difference is that in an agentic loop, a model decides when the work is done.

Brandom cited Boris Cherny, the creator of Claude Code, who told Meta's @Scale conference that loops are real and compared their arrival to the earlier shift from hand-written code to agents. Cherny described running one agent that constantly looks for ways to improve a codebase's architecture and another that hunts for duplicated abstractions to merge. Both submit pull requests and run indefinitely, because the code keeps changing.

The article also described a popular technique called the Ralph Loop, which periodically summarizes what a model has done and asks whether the goal has been met, sending it back to work if not. The method is meant to stop models from losing track during long runs. OpenAI researcher Noam Brown has observed that current models can solve most problems given enough compute, Brandom wrote, which makes a loop a way to keep applying compute until a problem is solved.

How are vendors building loops into their products?

The tooling is arriving quickly. On June 23, Chinese tech outlet AITNT published a detailed summary of a new OpenAI guide on using Codex for long-running work. According to that summary, the guide describes persistent threads that accumulate context and decisions, a memory store for ongoing work, and thread automations that wake an agent on a schedule and can keep running until a condition is met.

The examples are telling because they are not about writing code. One is a chief-of-staff loop that checks Slack and email on a schedule, flags important messages and drafts replies, while the user decides what gets sent. Another checks every five minutes whether a human agent has joined a refund conversation, then prepares the next reply within set limits. The guide frames the cycle as context, tools, memory, recurrence and review.

Google is moving in a similar direction. Its Interactions API, which became generally available on June 22, lets developers flag any call to run in the background on Google's servers, Google said.

Why do loops change the cost conversation?

Loops consume tokens far faster than a chatbot answering questions, and because they run continuously, TechCrunch noted, there is no natural ceiling on spending. Brandom pointed out that model providers sell tokens, so they benefit from heavier use, while customers may find the bills steep.

That is a familiar warning for anyone who has sat through AI Connect sessions on AI costs. Practitioners have described cloud spending that drifts for months before anyone notices. A loop that runs every five minutes, every day, across hundreds of employees, turns that risk into a structural one. Finance and platform teams will need per-loop budgets, alerts and kill switches, the same disciplines covered in AI Connect's session on AI ROI and cloud costs.

Who is supervising the agents?

The bigger question is oversight. A one-off agent request has a human at both ends. A loop may run for days with nobody reading each step. TechCrunch flagged token spending and drift as issues users must manage, and said trusting agents to run unsupervised is a large step.

AI Connect speakers have been describing this shift for months. In a AI Connect webinar on moving from human in the loop to agent ops, Adrish Sannyasi, VP, Customer Solutions and Delivery at Rhino Federated Computing, described people moving "from being an operator to more of a supervisor," with models proceeding on confident decisions and sending a short list of exceptions to humans.

That model only works if the exceptions surface reliably. Snowflake's Sam Mittal, VP of AI Solutions, warned in a AI Connect webinar on AI agent observability that "production AI that works on Monday can silently degrade by Thursday." For a loop, silent degradation does not just mean one bad answer. It can mean days of compounding work built on a wrong assumption.

How should enterprises design loops that last?

The practitioners who have built production agents offer a few consistent lessons.

Keep each loop narrow. "Don't build super agents, build focused agents," SAP Digital Solution Advisor Shubham Trivedi said during a AI Connect webinar on scaling AI architecture. A loop with one job, such as merging duplicate code or triaging one queue, is easier to measure and easier to stop.

Plan for failure inside the loop. In the same session, Skyflow's Joe McCarron, Head of Developer Advocacy, said an agent that gets error information back can steer itself to other tools. "You shouldn't have to fail the loop if the single tool fails," he said. Resilience matters more when nobody is watching in real time.

Make the goal testable. The OpenAI guide, as summarized by AITNT, argues that clear, checkable goals work better than vague plans, using the example of porting a software library while keeping its original unit tests as the acceptance standard. Enterprises can apply the same rule outside code: define what done looks like in terms a machine can check.

Write down the boundaries. The refund example works because the agent's actions are limited when no human is present. Every loop should have documented permissions, a spending cap and an owner who reviews its output on a set schedule.

What comes next

Loops are early, and most enterprises are still getting single agents into production. But the tools to run work continuously are now shipping from the largest providers, and employees will find them. The organizations that benefit will be the ones that treat a loop like a new team member: a defined role, a budget, a manager and regular performance reviews. The ones that do not may discover their first loop on an invoice.

Frequently asked questions

What is an AI agent loop?

An AI agent loop is an agent that keeps working in the background, often on a schedule, until a model judges that a goal has been met.

Why are AI agent loops expensive?

They consume tokens continuously rather than per question, so without caps there is no natural limit on spending.

How should enterprises govern agent loops?

Give each loop one narrow job, a testable goal, documented permissions, a spending cap and an owner who reviews its output regularly.

Filed under: Agentic AI, Observability & Data Quality, AI Strategy & ROI