Workflows

What a workflow is

A workflow is a sequence of agent tasks arranged as a directed acyclic graph (DAG). Each node is a task: run a script, process data, train a model, or send a notification. Edges define dependencies: task B waits for task A to finish before starting.

Workflows are for repeatable, multi-step pipelines. Things you'd otherwise do manually in the same order every time.

The DAG builder

Open workflows from the top menu or press Cmd+Shift+W. The builder is a visual canvas where you:

1. Add nodes. Click the + button or drag from the node palette
2. Connect nodes. Drag from one node's output port to another's input port
3. Configure nodes. Click a node to set its parameters

The graph validates as you build. Meshia flags cycles (not allowed) and missing connections.

Node types

| Node | What it does |

|---|---|

| Script | Runs a shell command or Python script |

| Agent task | Sends a prompt to an agent and waits for completion |

| Condition | Branches the workflow based on a metric or file check |

| Data | Downloads, preprocesses, or splits a dataset |

| Checkpoint | Saves or loads model checkpoints |

| Notification | Sends an email or webhook when reached |

Templates

Meshia includes starter templates for common workflows:

• Fine-tune and evaluate. Data prep, training, evaluation, results summary
• Hyperparameter sweep. Parallel training runs with different configs, comparison
• Nightly retrain. Pull new data, retrain, evaluate, deploy if improved

Click Templates in the workflow builder to start from one.

Scheduling

Workflows can run on a schedule. Click Schedule on a saved workflow and pick a frequency:

• One-time (run now or at a specific time)
• Daily, weekly, or custom cron expression

Scheduled workflows spin up a GPU pod automatically, run the DAG, and shut down when complete. You get a notification with results.