Skip to main content

Run a Workflow

Media placeholders

Before publishing, capture Run Workflow, an active execution, a human-input request, Rerun from Saved State, and final artifacts. Redact repository paths and external identifiers.

Running a workflow creates an execution you can follow from input to final output. Review the workflow's side effects before starting it; a valid graph can still use repositories, credentials, MCP tools, or external systems.

Before you run

Confirm that:

  • the workflow is marked valid;
  • you understand its required inputs;
  • referenced agents, providers, and MCP servers are available;
  • the selected repository and branch/workspace strategy are correct;
  • any required external credentials are configured;
  • high-impact actions have an appropriate human checkpoint.

Start the workflow

  1. Open Workflows.
  2. Find the workflow. You can run it from the list or open its detail page and select Run workflow.
  3. Complete the Run Workflow form.
  4. Recheck repository paths, ticket or pull-request identifiers, branch names, and deployment targets.
  5. Select Run Workflow.
  6. Allen opens the new execution so you can monitor it.

The Run action is disabled for a workflow with validation errors.

Fill the generated form

The form comes from the workflow's saved input schema. It can render text fields, textareas, checkboxes, numbers, selects, and repository pickers.

  • Required fields are marked and block submission when empty.
  • Optional blank values are omitted.
  • Boolean and number values are converted to their declared types.
  • Defaults are prefilled when the workflow defines them.
  • A repository picker lists registered repositories and also allows a manual path.

The repository picker appears only for a matching path field when the workflow declares that it requires a repository. If the form is missing an expected field, edit the workflow input schema rather than inventing another input name.

Monitor the execution

The execution detail page is the source of truth for the run. Use it to inspect:

  • overall status and current progress;
  • node status and child agent executions;
  • logs, tool calls, and traces;
  • model usage and estimated cost;
  • repository/workspace and linked chat context;
  • context selected for the run;
  • saved artifacts and final output.

Do not treat a completed status alone as approval to ship. Review the output and any external changes the workflow made.

Respond to human input

When the status is waiting for input, read the question, summary, evidence, and available actions before responding. The execution does not continue until the required input is submitted.

If the requested decision is unclear or the evidence is insufficient, do not guess. Leave the run paused and investigate the linked logs, artifacts, or repository state first.

Cancel a run

Use Cancel when the target is wrong, the input is unsafe, the workflow is repeating unproductive work, or an external prerequisite has changed.

Cancellation stops future workflow progress, but it cannot undo side effects that already happened. Check for commits, tickets, messages, deployments, or child executions before assuming the system is unchanged.

Rerun from saved state

Allen writes a saved state after each successfully completed workflow node. Open Rerun from Saved State to inspect those checkpoints.

For a completed, failed, or cancelled run you can:

  • compare saved states;
  • add feedback for all agent nodes or selected agent nodes;
  • edit checkpoint state while the execution is not active;
  • Run from a checkpoint using the same execution lineage;
  • Fork a checkpoint into a new execution for comparison.

Prefer a checkpoint rerun when earlier work is still valid. Start a fresh run when the original input, repository state, or intended path has materially changed.

Review the result

Before using the result:

  1. Read the final response and any declared workflow outputs.
  2. Open saved artifacts and confirm they contain the expected evidence.
  3. Review linked repository or external-system changes.
  4. Confirm validation or test results rather than relying on an agent's claim that they passed.
  5. Record follow-up feedback if you plan to rerun from a checkpoint.

Troubleshooting

SymptomWhat to check
Run is disabledOpen the workflow, validate it, and resolve its errors.
Expected input is missingCheck the saved INPUT schema and repository requirement.
Repository is not listedRegister it under Repos, or use manual path entry only when the workflow expects that path.
Run waits indefinitelyCheck the current node, pending human input, child executions, logs, and provider/tool health.
Node failsInspect the node trace and logs before retrying; fix input, credentials, model, MCP, or workflow logic as indicated.
No saved states appearA checkpoint is written only after a node completes successfully.
Output exists only in logsUpdate the workflow or agent instructions to return named output or save an artifact.