RoutineOSBeta
Lifecycle

Done means live and tested. Not "the agent says it works".

Agentic workflows go wrong when nobody checks them. Every RoutineOS routine goes through the same eight stages, and each stage ends with something you can look at before the next one starts.

Get startedBuild your first routine
from the docs8 stages3 states
The eight stages

Agentic workflows, from Discover to Retire, with proof at each step

The RoutineOS skill takes every routine through these stages. The exit evidence is the thing that must exist before you move on. Once a routine is live, Operate and Improve repeat.

01

Discover

Interview, or read a brief you already have. This stage is optional.

Exit evidenceA written brief you accepted.
02

Design

Pick routines, the mix of code and model, store, gates, secrets pattern and cost. List the risks.

Exit evidenceA design you approved, risks listed.
03

Build

Scaffold (generate the starting files), write prompts and scripts, build the Desk.

Exit evidenceLint (the automatic rule check) passes; every external contact declared.
04

Verify

Dry checks, then a real report-only first run.

Exit evidenceA run record exists and you looked at the output.
05

Deploy

Show the permission screen, create the routine and triggers, load seed data.

Exit evidenceThe routine is listed on the account with the stated schedule.
06

Operate

Read run records, watch health and cost, relay approvals.

Exit evidenceHealth and cost visible on the Desk.
07

Improve

Change or add routines and integrations without losing data.

Exit evidenceChange applied, verified, summarised in plain words.
08

Retire

Pause or archive cleanly, remove secrets.

Exit evidenceNothing scheduled; secrets deleted.
Three states

Only live and tested counts as finished

A routine is always in one of three states. An agent saying a routine works is not evidence.

State 1Not live. It does not exist on the account, or it is not scheduled.
State 2Live, not yet tested. It is scheduled, but nobody has checked a first run.
State 3Live and tested. The first run happened and you confirmed the output is what you wanted.
Why the first run is report-onlyA first run is report-only: the prompt forbids sending, posting or deleting. So you can check the output before anything leaves.
Common pitfalls

Three common mistakes

  1. Local time in the schedule.Schedules are in UTC (world standard time), not your local time. Convert before you set one.
  2. Choosing a type by the output.Choose by who acts. If the routine ever sends, it is approval-gated or it carries a gate. See the four types.
  3. Calling it done after deploy.Deploy makes it live, not tested. Check the first report-only run yourself.

One more from the docs: a routine that reads private data, reads untrusted content and sends outside should be split into more than one routine. Read the threat check.

Build your first routine

Take a template to a first report-only run. Start from Jobs watch or browse all templates.

Get startedRead the Quickstart