Start here · Automation owner

Start as an automation owner

Hand recurring checks, digests and reports to an agent. Each automation runs on the gateway at a fixed interval or on request, keeps every run as a readable chat turn, and speaks up only when a run needs you.

Install, try it, make it yours

Install and set a model 20 min

curl -LsSf https://raw.githubusercontent.com/lpalbou/AbstractFramework/main/scripts/install.sh | sh

Finish the first-run guide with a model that fits (or a cloud provider). Automations run unattended, so pick a model that answers reliably; a tick takes as long as your workflow does.

Create your first automation 10 min

Open the Observer from the console’s Apps page. Choose Launch, switch to Automate, pick Gateway default, and write the task:

Check the free disk space on this computer and report it in one line.

Set Repeat every 30 minutes, Independent context, tools run without asking, then Create automation. Already chatting in the Assistant? Its clock button, Schedule this conversation…, creates the same kind of automation from the conversation.

The Observer's Edit automation form, prefilled with the title, task, a 30-minute interval, growing context and tool approval
The Observer’s automation form: task, interval, context mode and tool approval.

Read it like a chat 5 min

The Automations page shows each run as a trigger turn and an answer turn. Run now fires one immediately: the schedule does not move, it does not count toward a run limit, and it works while paused. Open run ledger shows every step the agent took.

Decide what deserves your attention 15 min

Results are quiet by default. You are notified when a run fails after its last retry, when it waits for you, or when the workflow’s output carries notify. To alert on a threshold, publish a small workflow in Flow that returns {response, notify}; the gateway decides from that field, never from the wording of the answer.

Manage it from anywhere 10 min

Pause, resume, run now, stop the current run, edit (title, interval, context), archive, or Discuss a result in a forked chat, from the Observer, the Assistant’s Automations tab or AbstractCode:

/automations          # in the AbstractCode terminal client
/schedule <task>      # create one from the current workflow

Then make it yours

TRY 1

Market watch (growing)

"Search the web for the current price of <a share> and report it with the change since the previous check." Every 5 minutes, growing context: each tick sees the previous ones.

Growing context

TRY 2

Memory monitor (independent, notify)

A workflow that runs memory_pressure (a macOS command) and returns notify: true only under a threshold. Every 2 minutes; silent until it matters.

Worked examples

TRY 3

Weekly report you can question

Every 7 days, summarize the week’s changes in a folder. When a report surprises you, Discuss it: the new chat sees the automation’s history and can read its files, without changing them.

Discuss a result

Worth knowing

Intervals are fixed and in UTC. "Every 24 hours" means 24 hours after the previous tick, not "daily at 08:00"; there is no cron expression or time zone. Write weeks as 7d.
Creating an automation approves its tools. With the default approval, the framework's tools run without asking on every tick. Choose Ask each time if every tool batch should wait for you. Third-party MCP tools always ask.
Busy models slow ticks down. At most one run of an automation runs at a time; ticks that come due meanwhile are coalesced. Several automations sharing one local model queue behind each other. Growing automations replay the newest whole turns up to 50,000 tokens; ask for a short running summary in each answer when something must be remembered longer.