An always-on agent is an AI assistant that runs on a provider's servers instead of your own computer, and starts by itself: at a set time, or when something happens, such as a new order. Your laptop can stay closed. A webshop could, for example, have one sort the night's customer emails every morning at seven and post a summary in the team chat.
How does an always-on agent start?
It starts from a trigger instead of a person typing a prompt. Anthropic's routines for Claude Code, a research preview since 14 April 2026, support three kinds (Anthropic):
- A schedule, such as hourly, nightly or weekly.
/schedule daily PR review at 9amin the terminal creates one from a plain description. - An API call: each routine has its own endpoint and token, so a monitoring tool can start it with an HTTP POST.
- 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.
Before routines, teams managed their own cron jobs, infrastructure and tooling such as MCP servers for this, Anthropic writes. Agents built on the Claude Platform get schedules through Claude Managed Agents. "Always-on" is about availability: the agent is ready at any hour, and each trigger starts a separate run.
Where does it run?
In a sandbox: an isolated computer in the cloud with its own files and tools. In the routines documentation of April 2026, each run is a full Claude Code session on infrastructure that Anthropic manages. It clones the repository from the default branch, pushes changes to claude/ branches and runs without approval prompts.
OpenAI's Agents SDK update of 15 April 2026 gives developers who build their own AI agents native sandbox execution, in their own sandbox or at providers such as Cloudflare, E2B, Modal and Vercel. The harness, the code that runs the agent loop, stays outside the sandbox, which keeps credentials away from the code the model writes.
How does it remember anything between runs?
A cloud run starts clean, so memory has to live outside it. For routines, that is the repository. OpenAI's SDK stores the agent's state outside the container, so a run can continue from its last checkpoint in a fresh container when the old one fails or expires. LangAlpha, an open-source financial research agent that ginlix-ai started on GitHub in January 2026, keeps an agent.md file in each workspace with goals, key findings and an index of threads and files; its README of April 2026 says a middleware layer adds that file to every model call.
Where does it go wrong?
- It acts as you. Commits and pull requests from a routine carry your GitHub user, and Slack or Linear messages go out under your linked accounts. The docs advise giving each routine only the repositories, network access and connectors it needs.
- Secrets sit in a process that nobody watches. Routines take API keys as environment variables; LangAlpha masks known secret values in tool output before the model sees them. OpenAI advises assuming that prompt injection and data exfiltration will be attempted.
- Runs cost usage while nobody watches. At launch, routines counted against the subscription like interactive sessions, with daily caps of 5 runs on Pro, 15 on Max and 25 on Team and Enterprise.
- Some tasks stay on your own machine. Work behind a login is easier in the user's own browser: rtrvr.ai's AI Subroutines (16 April 2026) record a task in a Chrome extension and replay it inside the web page, where the user's cookies and CSRF tokens come along.
Example
Anthropic's announcement describes a routine that runs every night at 2am: pull the top bug from Linear, attempt a fix and open a draft pull request. LangAlpha's automations start on a cron schedule or when a stock hits a price condition. The April 2026 launches are compared in Can always-on agents replace local AI workflows?