AI builders embrace always-on agent swarms for coding and research

AI builders are no longer talking about agents as occasional helpers. They are describing background systems that keep running, hand work to other agents, and keep iterating with less human interruption. That shift is moving AI from one-off chat into persistent software, and it is changing how code gets written, how research gets done, and who carries the risk when something goes wrong.

What a loop actually is

A loop is not just a longer prompt thread. In the way builders are now using the term, it is software authorized to keep working continuously in the background, checking its own outputs, calling more agents, and returning to the task until it hits a stopping point. Boris Cherny, the Claude Code creator at Anthropic, captured the mood at Meta’s @Scale conference in Menlo Park, California, when he said, “loops are for real.”

Cherny described the trajectory as a clear progression: humans write source code by hand, agents write code, and then agents prompt other agents that write the code. He also described his own workflow as running multiple background agents, including one that looks for architecture improvements and another that searches for duplicated abstractions to unify. The practical meaning is simple: the agent is no longer waiting for a person to press send after every step.

How the model works in coding and research

Anthropic has already put this logic into its Research feature. The company said the system uses multiple Claude agents, with one agent planning the research process and parallel subagents searching at the same time. Anthropic has said that setup is especially useful for open-ended research tasks and breadth-first questions that need many independent directions explored at once.

That is a useful distinction. A single assistant is good at quick answers, but a multi-agent loop is built for situations where the first pass is rarely the last pass. In practice, the system can branch, compare, and revisit, which makes it better suited to sprawling questions, large codebases, and work that benefits from parallel exploration rather than linear conversation.

Why the industry is leaning into always-on work

OpenAI is making a similar bet with Codex. The company says the product is designed for multi-agent workflows and that its cloud environments let agents work in parallel across projects. It also says its automations handle always-on background work such as issue triage, alert monitoring, and CI/CD.

The business case is becoming clearer because the work is changing shape. OpenAI said on June 11, 2026 that Codex has more than 5 million weekly users, up 400% from earlier in the year, and that its most valuable work is increasingly unfolding over hours or days rather than minutes. The company also said non-developers now make up about 20% of users and are growing more than three times as fast as developers, which shows that these systems are moving well beyond traditional software teams.

What the infrastructure has to guarantee

The jump from experiments to production is where security starts to matter. OpenAI said its June 11, 2026 acquisition of Ona was meant to expand Codex with secure, customer-controlled cloud infrastructure for long-running agents across software and knowledge work. In that model, agents can keep working even when laptops are closed, but they do so inside controlled environments with scoped credentials, logging, and review.

That infrastructure matters because an always-on agent changes the basic trust equation. When work runs for hours or days without a person watching every step, access controls have to be tighter, audit trails have to be clearer, and review has to happen before a bad recommendation becomes a deployed change. The more persistence you give the system, the more important it becomes to limit what it can touch.

Where these systems are most likely to fail

The appeal of loops is also their biggest weakness: they can keep going. A persistent system can keep chasing the wrong objective, doubling down on a flawed assumption, or producing polished but redundant work that looks productive without being correct. That is especially true when multiple agents are working in parallel, because parallelism can create conflicting paths, duplicated effort, and more output than any person can realistically inspect.

OpenAI’s own SWE-bench Verified framing helps explain why. The company described coding-agent evaluation as giving a repository and issue description to a model and asking for a patch, and said evaluation itself remains an active research problem. As of August 5, 2024, top-scoring agents were at 20% on SWE-bench and 43% on SWE-bench Lite, a reminder that even strong systems still struggle on realistic engineering work.

Why cost and accountability are changing with the tooling

Loops also change the economics of AI use. A system that works for hours or days is not just another chat session, it is a compute commitment, and that creates a new budget line for organizations already under pressure to show clear returns. If agents are continuously scanning code, monitoring alerts, or exploring research branches, the cost of letting them run can rise quickly even when the human effort falls.

The legal and organizational question is sharper still: who is responsible when an always-on agent makes a bad change, mishandles data, or acts outside the bounds a team intended? The answer cannot be the model itself. That accountability lands with the company that deployed the system, the team that granted access, and the managers who decided the safeguards were enough.

What this shift really means

The move toward loops is not just a product trend. It is an argument that software should not stop at answering questions, but should keep working like a tireless junior operator that can search, compare, monitor, and revise in the background. Anthropic’s multi-agent research systems and OpenAI’s Codex push both point in the same direction: more autonomy, more parallelism, and more persistence.

That is why the language around agents is changing so quickly. The industry is no longer selling a smarter chatbot, but a layer of software that can stay in motion while people step away. The gains are real, but so are the new burdens, and the teams that adopt these systems will have to govern them with the same seriousness they reserve for any other high-stakes infrastructure.