Question: Enforcing sequences, loops, and steps

[verdverm]

Hypothetical workflow, implement some feature as an agent team.

1. Understand problem and context
2. Formulate plan
3. Implement plan (loop)
   1. Delegate work
   2. Review changes
   3. iterate / break
4. Final review (each step must run)
   1. Create summary / PR
   2. Code quality / conventions
   3. Security review

---

From what I gather

• the philosophy of scion is not to include workflow constructs like ADK (https://adk.dev/agents/workflow-agents/)
• instead, use markdown, the messaging system, and agents with access to the scion CLI
• ADK support means we may be able to build (4) and make that available to the overall scion task effort

My main concern around not having the workflow constructs is that markdown only agents are reliably unreliable.

Thoughts and insights?

[ptone]

You are correct in your interpretation.

You might be interested in the discussion on #176

It is not that I believe that more deterministic flows have no value or purpose, it is:

1. Scion is specifically focused on the range of problems where having LLMs orchestrate and navigate the workflow - while the models may be unreliable today, my experience in this space spans 3+ years to when models could not reliably complete one line of code.

2. Deterministic workflow wrappers can always call the scion CLI to maintain such a deterministic flow outside of scion

There are lifecycle hooks in agents (it is more of a latent capability, than a documented feature) that could be used to better integrate with an adjacent workflow system.

I'd be happy to hear more about how and where Scion would need to be modified to better support N different deterministic workflow system. The goal should be to find the right abstractions that allow effective loose coupling.