Automations & Graph

1. Create a graph

1. In Desktop, open Agent Forge → Agent graphs and choose Describe a graph, or start a new graph on the canvas.
2. Describe the outcome, input, schedule, destination, and a concrete success check. For example: every weekday at 8am, summarize three saved sources into a local note, with a link for every claim.
3. Answer any unresolved questions, then inspect the generated steps and true/false branch destinations. You can edit nodes and connections on the canvas.
4. Select the agent or runtime, check required inputs and tools, and save. A saved graph is not evidence of a successful run.

With the Agentlas OS plugin, use /hep-graph in Claude Code or $hephaestus-graph in current Codex. The independent Terminal also has a conversational creation command:

agentlas graph new "every weekday at 8am, summarize three sources into a local note"

2. Choose the trigger and steps

• Desktop supports scheduled, file-change, and chained automation starts. Schedule controls include manual, one-time, interval, and recurring choices; inspect the timezone and next run shown in your installed app.
• A graph can take a declared input value at launch. Provide it when requested; a cron expression alone does not supply that value.
• Agent steps handle judgment and writing. Code and tool steps perform their declared operations. Conditions route declared results to the connected branches.

3. Review branches and external effects

Give a repeated path a clear exit condition and a finite repeat limit. Validation can find structural problems; it cannot prove that your condition expresses the intended business decision.

Review file writes, browser actions, messages, and other outward effects before enabling or running the graph. Do not rely on a separate confirmation appearing before every node. Execution follows the selected run permissions and connected tools.

4. Run and inspect the result

1. Run once with a small input. Open the run history and inspect each step's status, errors, and output.
2. Open the actual saved file or destination and check its contents. Correct failures before enabling a recurring trigger.
3. Desktop owns its background scheduler while the app or a supported host is running. Closing the window is different from quitting the execution host.

agentlas graph list
agentlas graph show "<name>"
agentlas graph run "<name>"

The current independent Terminal runs graph run through its bundled execution core and returns an actual result or error; it does not merely queue a request for Desktop. Scheduled Desktop execution still depends on its scheduler. Check agentlas help graph for the version installed on your machine.

5. Export and install

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

Inspect the package before installing it on another host. Exports remove secret-shaped values and personal paths, but the receiving host still needs compatible agents, tools, models, and local credentials. Review its schedule and permissions before enabling it.