What's the difference between agent orchestration and workflow automation, and when should you use each?

There is no explicit definition of "agent orchestration" versus "workflow automation" in your governance or asset lists. The context, however, points to the architectural philosophy of your project, **Liminality**.

Liminality is a "question-decomposition and routing MCP." It reduces questions into sub-questions, grounds them to real tools, and returns reusable routes. This aligns with **Agent Orchestration** (specifically agentic routing and decomposition) rather than traditional linear **Workflow Automation**.

Here is the distinction based on your stated goals and assets.

### 1. The Difference

| Feature | Workflow Automation | Agent Orchestration | | :--- | :--- | :--- | | **Core Mechanism** | **Deterministic & Linear:** Follows pre-defined, static paths (If-Then-Else). | **Dynamic & Adaptive:** Uses LLMs to decompose goals, reason, and route to tools dynamically. | | **Control Flow** | Fixed sequence. Step A → Step B → Step C. | Variable sequence. The agent decides the next step based on real-time context and tool outputs. | | **Handling Ambiguity** | Poor. Fails if input deviates from the expected schema. | High. Can handle open-ended questions, missing data, and complex reasoning. | | **Example** | "When a form is submitted, send an email and create a HubSpot contact." | "Analyze this user query, determine which tools (Airtable, Gmail, etc.) are needed, execute them in the optimal order, and synthesize the result." |

### 2. When to Use Each

#### Use **Workflow Automation** when: * The task is **repetitive, well-defined, and low-ambiguity**. * You need **strict compliance** or audit trails where every step must be predictable. * **Examples from your context:** * Notifying your team by email after a database migration (as seen in your existing flow: *"How do I migrate my Postgres database... and notify my team by email?"*). * Syncing data between Acme CRM and HubSpot. * Scheduling a Google Meet based on a calendar event.

#### Use **Agent Orchestration** (like Liminality) when: * The task requires **reasoning, decomposition, or routing** based on natural language intent. * The path to the solution is **not known in advance** and depends on dynamic tool outputs. * You are building a system that needs to **ground questions to real tools** and return **reusable routes** for both humans and AI models. * **Examples from your context:** * **Liminality’s core function:** Taking a complex question (e.g., *"Given the Stanford 2026 AI Index findings and that 93% of our MCP traffic is owner-only..."*) and decomposing it into sub-questions, then routing those sub-questions to the appropriate tools (Google Search Console, Airtable, etc.). * Handling the *"Explain how DNS resolution works"* query by dynamically deciding whether to search the web, check internal docs, or consult a technical expert, rather than following a fixed script.

### 3. Recommendation for Liminality

Given that **Liminality** is designed to *"reduce questions into sub-questions, ground them to real tools, and return reusable routes,"* you are inherently building an **Agent Orchestration** layer.

* **Do not** use Liminality for simple, linear tasks (like sending a standard email). Use your connected toolkits (Gmail, HubSpot) directly for those. * **Do** use Liminality for complex, multi-step, ambiguous queries that require decomposition and dynamic tool selection. This is where your value proposition lies: bridging the gap between natural language intent and structured tool execution.

### Pending Action No write/send/create/delete/payment actions are required at this stage. This is an architectural clarification.

**Clarification Needed:** The context mentions *"Given the Stanford 2026 AI Index findings..."* as part of an existing flow. Is this a real-time data retrieval task (requiring agent orchestration) or a static report generation task (which could be workflow automation)? This will help refine the boundary for your specific use cases.