---

name: agents-workflows description: "Guides SuperAGI agent creation, workflow selection, prompt parsing, task queues, scheduling, and Celery execution loops." disable-model-invocation: true metadata: disco-role: operating license: MIT

SuperAGI Agents and Workflows

Use this sub-skill when the task is about how SuperAGI agents are prompted, selected, queued, scheduled, executed, paused, resumed, or advanced through workflow steps.

Read First

• references/workflows.md for workflow types, seed names, and execution flow.
• references/prompt-and-output-parsing.md for prompt builders, parser signatures, and output-shape expectations.
• references/execution-loop.md for the agent, job, and Celery loop from config to execution step.
• references/troubleshooting.md for parser, tool-selection, queue, permission, and scheduling failures.
• scripts/validate_agent_payload.py for a safe structural check of agent/run payloads.

Core Concepts

• Goal-based workflow: a single-step prompt loop that keeps selecting the next tool until completion.
• Dynamic task workflow: a task-analysis and task-prioritization loop that revisits the task queue.
• Fixed task workflow: a task queue that follows a fixed schedule of initialization and queue steps.
• Tool workflow steps: tool steps, wait steps, task-queue steps, and permission-gated steps are all represented in the workflow models.
• Execution loop: superagi.worker and superagi.jobs.agent_executor coordinate scheduled agents, waiting workflows, resource summaries, and webhook callbacks.

Typical Questions This Sub-skill Answers

• Which workflow name should I assign to an agent template?
• Why did the agent choose the wrong tool or produce invalid JSON?
• Why is a task queue empty, repeated, or not advancing?
• How do permission checkpoints and wait steps change execution?
• Which Celery task or workflow seed owns a given behavior?

Safe Workflow

1. Route by workflow family: agent templates and workflow names belong here, not in API or toolkit sub-skills.
2. Read prompt and parser references before changing or validating outputs.
3. If the user provides a payload, use the validator script for a structural check before thinking about execution.
4. For real execution, remember that workflow code depends on Redis, a database, and in many cases an LLM provider. Only run live execution when authorized.

Boundary Notes

• HTTP route and model schema questions belong to api-service.
• Tool implementation and toolkit registration belong to toolkits-integrations.
• Provider, resource, and vector-store setup that a workflow merely consumes belongs to models-resources-vector.