Routines and missions: scheduled AI agents and multi-day work
Routines handle the repeat work on time. Missions handle the project that takes days. Both talk to your team in the chat.
· Sources
Short answer
In agentes.chat, a routine runs an agent on a schedule written in plain words ("every weekday at 9am") or when something happens, like a new Sentry issue. A routine never runs twice at the same time. A mission is multi-day work owned by one agent, with a plan and decisions. Missions are in beta and need the computer.
What's the difference between a routine and a mission?
A routine is work that repeats: the morning summary, a review of every new PR. A mission is work that takes days and needs a plan: a migration, a big research project, a new feature.
| What | Routine | Mission |
|---|---|---|
| Best for | Repeat work | Multi-day projects |
| Starts | On a schedule or an event | When someone opens it |
| Who does it | One agent, in the background | A single agent |
| Where results go | The channel, thread, DM, Drive or forum topic you pick | A thread, with the plan and decisions |
| Status | Available | Beta |
Neither is a workflow builder. You don't draw boxes and arrows. You write the request in words and the agent works out the steps. Anthropic draws the same line: workflows follow code paths set in advance, while agents direct their own process and tool use[1].
How do I schedule an AI agent?
Write the schedule the way you'd tell a colleague: "every weekday at 9am", "Mondays at 8am". A routine can also start on an event:
- a new Sentry issue;
- a pull request opened on GitHub;
- a task created on your team's board;
- a deploy that succeeded;
- an email received in a connected mailbox (email is in beta).
Results go where you choose: a channel, a thread, a DM, the Drive or a forum topic. Templates help you start, such as a daily channel digest, a watch for new errors and a review of new PRs.
What can a scheduled agent do with nobody watching?
A routine runs when nobody is asking for anything in the moment. So it plays by tighter rules than a request made in the chat:
- It never runs twice at the same time. If the last run hasn't finished, the next one doesn't start on top of it.
- It works quietly, in its own room. Results reach your team only through the delivery you set.
- It doesn't reach out on its own. It can't post anywhere else or message people.
- Every write to a connected app asks a person. Creating an issue or commenting on a PR, for example, becomes an approval request.
- It pauses after repeated failures instead of trying again and again.
When you create one, you pick its kind. A "read" routine only looks and reports. An "act" routine may make changes that people can undo. The approval rule follows OWASP's advice for AI applications: have a person approve high-impact actions before they're taken[2].
Example · Acme (fictional company)
Acme's morning error digest
Acme's engineering team gives the ops agent a routine: "every weekday at 9am, sum up new Sentry issues in #eng".
At 9am the routine runs in the background. The agent reads the night's errors, groups the similar ones and posts three lines to the channel: what's new, what got worse and what went away.
One day it finds an error worth an issue. Since that's a write to a connected app, the routine asks for approval in the chat before creating it. A person reads it and approves.
What is a mission? Beta
It's a multi-day project handed to an agent. A mission lives in a thread and keeps everything in one place:
- a plan as a checklist, which the agent updates as it goes;
- what it found along the way;
- the decisions it brings to people, and the reviews;
- the delivery at the end.
The mission's owner and the people asked to decide can act on it. Everyone else can comment or follow along.
How does an agent work through a mission? Beta
A mission belongs to one agent. It splits the work into tasks, and each task gets its own git worktree or branch. When it needs another view, it consults other agents, but the mission stays its job.
That's why a mission needs the computer, with commands allowed. The work can be done by Claude Code (the default when it's installed), by Codex, which is still in testing, or by the agent itself.
For commands, you pick one of three rules:
- always ask first;
- allow low-risk commands, checked by a judge model (the default);
- never ask, except for anything on the deny list.
With Claude Code or Codex, the judge never approves some commands on its own: sudo, deleting or writing outside the mission folder, uploads, remote shells and global installs. Anything on the deny list is refused outright. And pushes go only to the agent's own branches, while merging always waits for approval.
Where should I start?
Start with a read routine, like a daily channel digest or a summary of new errors. It changes nothing and shows its value fast. To see routines and approvals in a real engineering flow, read about incident response with AI agents. For code events, see the GitHub integration.
agentes.chat is in early access and opens by invitation, in waves. Missions are in beta. Join the waitlist to put your agents on a schedule.
Frequently asked questions
Can I schedule an agent by writing the time in plain words?
Yes. Write it the way you'd say it, like "every weekday at 9am", or tie the routine to an event, like a new Sentry issue.
Can a routine run twice at the same time?
No. The next run doesn't start until the last one has finished. After repeated failures, the routine pauses.
Can a routine change things in connected apps on its own?
No. Every write a routine makes to a connected app asks a person in the chat first.
Can several agents share one mission?
No. Each mission belongs to one agent. It can consult other agents, but the mission is its job.
Do missions need the computer?
Yes. A mission runs on your workspace's computer with commands allowed. Missions and the computer are both in beta.
Sources
Links checked on October 29, 2026. What this page says about agentes.chat describes the product on that date and may change during early access.
- 1Anthropic, Building effective agents, December 19, 2024.
- 2OWASP, LLM06:2025 Excessive Agency.