Guides

Automations & Graph

Graph is Agentlas's automation builder: describe what you want in a sentence, answer a few questions, and get a runnable step graph — without drag-and-drop node wiring and without an unbounded loop or a silently-wrong branch ever being possible to build.

INFO · Three surfaces, two different jobs

You create a graph by talking to it — from Claude Code/Codex (/hep-graph) or the Terminal (agentlas graph new). You run and edit a graph from Agentlas Desktop. This split is deliberate, not a limitation: creation is a conversation, execution is something a scheduler needs to own reliably.

Creating a graph

agentlas graph new "every weekday at 8am, pull 3 blog topics and save them to notes"

You describe the outcome in plain language, in whatever language you're already typing in.
It asks up to 6 rounds of questions, at most 3 at a time — only for things it genuinely cannot decide for you (an ambiguous time, whether a step should repeat, which service you mean). If you say "you decide," it picks the safest option and tells you what it picked.
The model only ever produces a blueprint (name, goal, trigger, steps, branches) — the actual runnable graph is assembled by deterministic code, not written by the model node-by-node. That's what makes an unbounded loop or an undeclared branch structurally impossible to create, not just discouraged.
Before saving, it prints the built step tree and asks you to eyeball the branch directions — direction can't be self-verified by the system, so a human has to glance at it once.
A new graph is always saved switched off. You turn it on deliberately.

What a graph is made of

Trigger

cron (a schedule) or input (runs only when you supply a value). Nothing else — no open-ended trigger types.

Agent step

Judgment, writing, summarizing, deciding — the default step kind.

Code step

Exact computation a chat model would get wrong (date math, precise data-shaping), run isolated.

Condition / branch

Splits the path; direction is confirmed by you at creation time, never inferred silently later.

Mutation vs. read

Any step that leaves the machine or changes a file is flagged as a mutation and locked to ask for confirmation before it runs, unless you explicitly approve it up front.

Loops have to end

A branch can repeat back to an earlier step, but only with an explicit cap (1–20 repeats) and only by branching on a declared pass/fail check — never on raw text content. An automation that could loop forever with no one watching it is exactly what this is built to prevent.

Running and editing

agentlas graph run "<name>" doesn't execute the graph itself — it marks the run as requested, and Desktop's scheduler picks it up (within about a minute if Desktop is open, or at next launch). The Terminal never claims to have run it.
agentlas graph list / show "<name>" — see your saved graphs and their step tree without opening Desktop.
Editing steps, enabling/disabling, and the visual canvas all live in Agentlas Desktop, under Automations.

agentlas graph export "<name>" my-graph.agentgraph
agentlas graph inspect my-graph.agentgraph
agentlas graph install my-graph.agentgraph

An exported .agentgraph file has any secret-shaped values and personal file paths stripped before it's written, and a hardcoded model choice is softened to a general tier hint so it adapts to whoever installs it.

Open Automations in Desktop →