What's the difference between deterministic agent planning and prompting, and when should you use each?

There is no mention of deterministic agent planning versus prompting in the provided text. The source material covers Liminality’s architecture and assets, its intended adoption, and current flows like DNS resolution and database migration.

I am answering based on general technical knowledge instead. This is external to the Liminality context.

### Deterministic Agent Planning vs. Prompting

Control flow and reliability define the difference.

#### 1. Prompting (Non-Deterministic / Probabilistic)

You send a natural language instruction to an LLM. The model generates a response using probabilistic weights.

* **Flexible:** It works for open-ended tasks, creative writing, summarization, or ambiguous queries. * **Unpredictable:** Outputs vary between runs. Temperature settings and model updates cause this drift. * **No Guaranteed Structure:** The LLM might ignore formatting constraints or skip steps if not forced.

Use prompting when: * The task is ill-defined or needs creativity. * You need natural language understanding or generation. * The cost of a minor error is low.

**Example:** "Summarize this email thread and suggest a reply."

#### 2. Deterministic Agent Planning (Structured / Algorithmic)

The agent uses predefined logic, a state machine, or code-based planner. It breaks a goal into steps. The LLM acts as a tool within this plan, perhaps to extract information. Code or a strict protocol controls the flow.

* **Predictable:** The same input follows the same logical path. This assumes tools behave deterministically. * **Reliable:** Errors are caught at specific steps. The agent can retry or fallback. * **Auditable:** You can trace exactly which step failed or succeeded.

Use deterministic planning when: * The task requires zero downtime or high reliability. Think database migrations or financial transactions. * Steps are well-defined but complex. * You need to guarantee specific actions happen in a specific order.

**Example:** The flow you already have: "How do I migrate my Postgres database to a new server with zero downtime and notify my team by email?" This is a deterministic plan. Step 1 (Migrate) leads to Step 2 (Verify), which leads to Step 3 (Email).

### Summary Table

| Feature | Prompting | Deterministic Agent Planning | | :--- | :--- | :--- | | **Control** | LLM decides the path | Code/Logic decides the path | | **Output** | Probabilistic | Deterministic (for the same state) | | **Best For** | Creativity, ambiguity, NLP | Execution, reliability, complex workflows | | **Risk** | Hallucination, inconsistency | Rigidity, inability to handle ambiguity |

### Relevance to Liminality

Your Liminality system decomposes questions into sub-questions and routes them to tools. It is a deterministic planning layer on top of probabilistic LLM capabilities. It ensures complex goals like the Postgres migration break down into executable, reliable steps. It avoids relying on a single prompt to just do it.

**Pending Action:**

Do you want me to create a new flow in Liminality that compares these two approaches for a specific use case in your stack? HubSpot or Airtable are examples.