Proposal: Add Optional “Compiled Plan” Mode for Deterministic Agent Execution

[yuer-dsl]

Hi SK team!

I’ve been experimenting with different agent execution patterns,
and noticed something about the current Agent Orchestration flow
(Planner + Kernel).

Right now the Planner dynamically generates a task/step sequence using LLM reasoning.
This is extremely flexible, but also inherently non-deterministic:

• For the same input, different runs may follow different execution paths
• Small context changes can alter the Planner’s reasoning
• This makes reproducibility and replay debugging harder
• Audit logs show the steps, but not a stable route that can be re-executed

This is expected given LLM-driven planning, but it made me wonder:

Idea: Add an optional compiled plan execution mode

The idea is simple:

1. The user expression/request is converted once into a static execution graph (a “compiled plan”).
2. The runtime scheduler executes the graph deterministically, step by step.
3. Same input + same plan → same route + identical audit structure.

This would not replace dynamic planning
(LLM planning is extremely powerful and should remain the default).
It would just be a supplemental mode for workflows that prefer:

• deterministic behavior
• reproducibility
• easier debugging
• stable audit trails
• high-assurance task execution

Example: Minimal compiled plan format

Here’s an extremely small example of what a static plan might look like:

{
  "plan_id": "sample-plan-001",
  "nodes": [
    {
      "id": "analyze",
      "type": "llm",
      "inputs": ["user_message"],
      "outputs": ["analysis"],
      "next": ["tool_select"]
    },
    {
      "id": "tool_select",
      "type": "router",
      "inputs": ["analysis"],
      "outputs": ["tool_id"],
      "next": ["tool_call"]
    },
    {
      "id": "tool_call",
      "type": "tool",
      "inputs": ["tool_id"],
      "outputs": ["tool_result"],
      "next": ["final_answer"]
    },
    {
      "id": "final_answer",
      "type": "llm",
      "inputs": ["analysis", "tool_result"],
      "outputs": ["answer"],
      "next": []
    }
  ]
}

This isn’t meant to define a formal schema—
just a simple illustration of what a static route graph could look like.

 Why this could be helpful

A static compiled plan mode may help with:

deterministic multi-step workflows

reproducible agent routing

consistent evaluation and testing

audit logging (route + input/output pairs are stable)

debugging (easy to replay the plan)

fallback / rollback logic

It also gives SK two complementary execution modes:

dynamic planning (default) → flexible, model-driven

compiled plan execution (optional) → stable, deterministic

 If helpful, I can provide a tiny PoC

If the team is interested,
I can share a very small (30–40 line) PoC of a deterministic scheduler
executing a compiled plan like the example above.

Happy to contribute if this aligns with the roadmap!

Thanks for reviewing!