Skip to content

Workflow lifecycle

Agentic workflows let you define secure, structured pipelines that connect AI models, retrieval tools, and task-specific logic—all within a governed, auditable environment. Whether you're using hosted LLMs, private inference backends, or vector search services, every workflow follows the same lifecycle:

  1. Draft
  2. Review
  3. Launch

Draft stage

In the draft stage, you define your workflow’s structure and logic.

Each workflow begins with a start point (where input arrives) and an end point (where the final result is returned). In between, you add and configure agents that perform distinct tasks—such as retrieving documents, calling a language model, or transforming output.

You can experiment freely during this stage. You’ll configure parameters like model type, temperature, or max tokens, and define context prompts that describe how each agent should behave. For example, a context prompt might instruct a model to act as a claims analyst, extract structured data, or summarize search results. Workflows remain editable until they’re submitted for review.

Review stage

Once your workflow is fully defined, you submit it for review.

The review process ensures that agent behaviors, prompts, and configurations meet your workspace’s security and governance requirements. Reviewers check that all required fields are present, that the agents are connected properly, and that the workflow does what it’s intended to do.

If any part of the workflow needs to change, reviewers can reject it and then the workflow creator returns it to draft for revision. Once approved by all required parties, it becomes eligible to launch.

Launch stage

After approval, your workflow is ready to launch.

Launching activates the workflow as a persistent, callable service. It enters a running state where it can receive input and return structured output using the Python SDK.

Once running, the workflow remains active until you pause or stop it. You can send multiple requests to a running workflow, each one returning a result in real time. These results might be answers, classifications, or structured decisions—depending on how you configured the workflow logic.

Because all executions take place inside an attested enclave, every response is compliant with your workspace’s data policies and protected from exposure—even during inference.

Ready to build your first agentic workflow? Get started.