Discussion: Progressive Reading DSL for Agent Workflow Management

[tracer-mohist]

Hi Codex team,

I have been exploring different approaches to manage AI agent workflows. I want to share an alternative design for discussion.

The Problem:

• Most workflow engines treat execution as jump to step N
• This assumes the agent has already read and understood all previous steps
• In practice, agents need to read progressively before executing

Our Approach (workflow-as-list):

• Workflows are lists of steps with dependency annotations
• Agent must read current step before advancing (enforced by steps_read tracking)
• Import system with caching for reusable workflow fragments
• State is externalized to files, not embedded in prompts

Key Design Decisions:

• Reader metaphor: workflow-as-list is a reader, not an executor
• Progressive reading: cannot skip unread steps
• Hash-only import naming for security (no path injection)

Repository: https://github.com/tracer-mohist/workflow-as-list

Question for Codex:

• How does Codex handle workflow state management?
• Do you track which steps the agent has processed?
• Curious about the tradeoffs you have made

Happy to discuss further or close if this is not the right venue.