Human-in-the-Loop Agent Workflows

Use a Stated publication as the boundary between autonomous AI work and human review, feedback or approval before the next workflow step.

2026-09-15 · 2 min read

A Stated artifact can act as the human-facing boundary in an agentic workflow. The agent performs work, publishes a result for review, and the surrounding application decides how a human response affects the next step.

Pattern

This is useful when a person needs to understand or respond to the agent's output before the workflow proceeds.

1. Produce the review artifact

Create or update a Stated Page or Experience with the decision-relevant output. Include the evidence and limitations the reviewer needs.

Do not publish internal chain-of-thought, credentials, hidden prompts or execution traces.

2. Add the appropriate response mechanism

Choose the smallest interaction that matches the review task. Examples include a poll for a small set of choices, a form for structured input, a rating for evaluation, a question for open feedback, or a voice interaction when spoken responses are appropriate.

3. Share with the reviewer

Apply the visibility and sharing rules required by the use case. Enterprise or client information should use access controls appropriate to its sensitivity rather than being made public by default.

4. Read the result

Use Stated's results and analytics interfaces to inspect responses and interactions. Your application or automation is responsible for deciding how those results map to subsequent actions.

Do not treat a page view as approval. Define the explicit response event your workflow requires.

5. Continue the workflow

Once your application has the required human signal, it can trigger the next agent task, update the artifact or route the case for further review.

For higher-risk actions, apply the authorization, audit and approval controls required by your organization in addition to the publication workflow.

Example: research prioritization

An agent publishes five research findings. The stakeholder selects the area that needs deeper investigation. The automation passes that selection into the next research task, and the same Stated Page is updated with the new findings.

Example: client feedback

An AI agency publishes a proposed strategy. The client submits structured feedback through the publication. The agency workflow uses the response to prepare the next iteration without exposing the underlying orchestration stack.

Persistent artifact

When the workflow represents one continuing business object, keep the same Stated artifact across cycles. Use stable identifiers and update operations to avoid duplicate publications.

See the Agentic Publishing and Persistent AI Artifacts documentation for the underlying architecture.