Commentary
Open appTry Commentary
Agent workflow guide

Human-in-the-loop should be a workflow—not a person watching a chat window.

A reliable human-in-the-loop workflow gives people a stable artifact and named decision stage. It does not depend on someone noticing the right message while an agent continues working.

  • Explicit artifact
  • Stable feedback
  • Revision verification
  • Human approval
Try the human-agent loopChoose an integration
human review workflow
Human approval
Implementation plan review
## Rollout and rollback

Deploy progressively, monitor the acceptance signal, and stop when the agreed threshold is crossed.

Structured revision context ready
Reference workflow

Turn human attention into a repeatable stage.

The artifact, feedback, revision, and approval state stay explicit from start to finish.

  1. 01

    The agent creates an explicit plan or review artifact.

    Start from an explicit artifact and named owner.

  2. 02

    The workflow opens a named human review stage.

    Keep the current state and unresolved context visible.

  3. 03

    Reviewers separate blocking and non-blocking decisions.

    Keep the current state and unresolved context visible.

  4. 04

    The agent retrieves unresolved feedback and revises the source.

    Keep the current state and unresolved context visible.

  5. 05

    Humans verify the new revision against accepted decisions.

    Keep the current state and unresolved context visible.

  6. 06

    Approval unlocks the next permissioned stage.

    The next stage begins only after verification.

The minimum reliable loop

Every implementation should define the artifact, reviewer, permission boundary, revision mechanism, and approval state.

Explicit artifactNamed review stageStable feedback recordsBlocking versus non-blocking decisionsRevision and verificationApproval stateAuditabilityAgent permission boundaries

Approval is not the same as execution permission

A reviewed plan may still require separate repository, deployment, or preview permissions. Keeping those boundaries explicit prevents a positive review from becoming an accidental authorization expansion.

A practical reference architecture

Use Draft Review for pre-Git plans, Markdown Review for source-backed documents, MCP or HTTP APIs for agent retrieval, and Live Preview Reviews only for applications that explicitly opt in.

Continue the workflow

Use the next page that matches the artifact or boundary you need to review.

Codex

Review a plan before implementation.

Read next

Claude Code

Add a durable review stage.

Read next

GitHub Copilot

Review plans and Markdown pull requests.

Read next

MCP & API

Automate structured handoff.

Read next

Security and access

Understand provider, preview, token, and human approval boundaries.

Review boundaries
Human review, made explicit

Put the workflow into practice.

Design an explicit, auditable human review stage for coding-agent plans, revisions, and approval.

Try the human-agent loopChoose an integration