Can always-on agents replace local AI workflows?

Between 13 and 15 April 2026, Cloudflare, Anthropic and OpenAI each shipped tooling for AI agents that run in the cloud, so the work no longer needs an open laptop. Such always-on agents start on a schedule, an API call or an event like a new pull request. They can take over jobs tied to a repository or an API. They did not replace everything local: Anthropic's Routines were a research preview with daily caps, and tasks behind a login still ran best in the user's own browser.

What is an always-on agent?

It is an AI agent that runs on a provider's infrastructure and starts from a trigger, without anyone watching. The usual local workflow is different: a developer types a prompt in a terminal or editor, waits for the result and steers. Anthropic describes what teams did before Routines: to automate Claude Code, they managed cron jobs, servers and extra tooling such as MCP servers themselves.

What did Anthropic ship with Routines?

On 14 April 2026 Anthropic introduced routines in Claude Code as a research preview. A routine is a prompt, one or more repositories and connectors to services like Slack or Linear, configured once. It starts in three ways:

• on a schedule, such as hourly, nightly or weekly
• through an API call: every routine gets its own endpoint and token, and a POST returns the URL of the new session
• on a GitHub event, such as a pull request or a release; Claude opens one session per pull request and keeps feeding it comments and CI failures

Anthropic's examples include a routine that pulls the top bug from Linear every night at 2am, attempts a fix and opens a draft pull request, and one that runs a team's own review checklist on every new pull request.

The documentation of April 2026 describes how a run works. Each run is a full Claude Code session in an isolated virtual machine managed by Anthropic. The repository is cloned at the start of every run, from the default branch, and Claude puts its changes on branches that start with claude/. There are no approval prompts during a run.

Routines were available on the Pro, Max, Team and Enterprise plans, with daily caps: 5 runs for Pro, 15 for Max and 25 for Team and Enterprise. Runs also draw down the same usage limits as interactive sessions.

What changed in OpenAI's Agents SDK?

On 15 April 2026 OpenAI released the next version of its Agents SDK, version 0.14.0 of the Python package. The harness, the code that runs the agent loop around the model, got configurable memory, file tools like those in Codex and standard support for MCP, skills, AGENTS.md instructions, a shell and file patches.

The larger change is native sandbox execution: the agent works in a controlled environment with its own files and tools. Developers can bring their own sandbox or use built-in support for Blaxel, Cloudflare, Daytona, E2B, Modal, Runloop and Vercel. OpenAI keeps the harness apart from the sandbox, which keeps credentials out of the environment where model-generated code runs. Because the agent's state lives outside the container, the SDK can take snapshots and continue from the last checkpoint in a fresh container when the old one fails or expires. The features are generally available at standard API prices, in Python first, with TypeScript planned.

What did Cloudflare add to its Agent Cloud?

On 13 April 2026 Cloudflare expanded its Agent Cloud, with the stated aim of moving agents "from experimental demos on local laptops" to workloads on its own network:

• Dynamic Workers, a runtime that starts in milliseconds to run a piece of AI-generated JavaScript and then disappears; Cloudflare claims 100 times the speed of containers at a fraction of the cost
• Artifacts, Git-compatible storage where agents can keep code and data
• Sandboxes, now generally available: a persistent, isolated Linux environment with a shell, a file system and background processes
• Think, a framework in Cloudflare's Agents SDK for long-running tasks with many steps

In the same announcement, OpenAI said that developers can now deploy agents in Agent Cloud that run on its GPT-5.4 model and Codex.

How do cloud agents remember anything between runs?

A cloud run starts clean, so whatever the agent needs to remember has to live outside it. With Routines, that is the repository itself: branches, pull requests and the files committed to it. OpenAI stores the agent's state outside the sandbox.

The open-source project LangAlpha (Apache 2.0), a financial research agent that calls itself a "vibe investing agent harness", shows a third pattern. Its README of April 2026 describes a dedicated cloud sandbox per workspace, with a file called agent.md at the root that holds the goals, key findings and an index of threads and files. The agent maintains that file, and LangAlpha adds it to every model call.

LangAlpha also uses programmatic tool calling: the agent writes and runs Python in the sandbox and passes only the result back to the model, instead of pouring raw market data into the context window. It can run subagents in parallel and has automations that start on a cron schedule or when a stock hits a price. During the beta, the price triggers worked only on the hosted platform.

Does everything move to the cloud?

Tasks behind a login are harder to move. On 16 April 2026, rtrvr.ai presented AI Subroutines, which record a browser task in its Chrome extension and replay it from inside the web page itself. Cookies, CSRF tokens and signed headers come along automatically, where a scraper running outside the browser has to rebuild all of that. A replay costs no tokens and gives the same result for the same input, according to rtrvr.ai. It names the limits as well: it records HTTP requests, so sites that work over WebSockets need slower actions in the page, and a Subroutine has to be recorded again when a site changes its API. This kind of deterministic web automation runs in the user's own browser, on the user's machine.

What are the risks of agents that run unattended?

• A routine acts as you. Commits and pull requests carry your GitHub user, and messages in Slack or tickets in Linear go out under your linked accounts. The docs advise limiting the repositories, network access and connectors to what the routine needs; the default environment only reaches an allowlist of domains.
• OpenAI advises designing agent systems on the assumption that prompt injection and attempts to exfiltrate data will happen, which is why it keeps credentials away from the code the model writes.
• Unattended runs cost usage: every routine run counts against the subscription, and during the preview GitHub events also had hourly caps per routine and per account.
• Long runs fill the context. In a July 2025 report, Chroma tested 18 models and found that their performance grew increasingly unreliable as the input got longer, an effect known as context rot.

Anthropic said in April 2026 that it planned to let routines start from more event sources than GitHub.