DEV Community

Alexander Ertli
Alexander Ertli

Posted on

When to Use OpenAI + Tools vs a Workflow Runtime

#ai #llm #saas #webdev

Modern “agentic AI” needs more than prompts—it needs architecture.

This guide shows when to stay inside OpenAI’s style tool ecosystem and when to move to a workflow runtime for observability, safety, and control.

💡 TL;DR

  • Use OpenAI + function calling or MCP when your AI just needs to answer a question. Maybe call one or two tools. All in one turn.
  • Use a workflow runtime when your AI must run multiple steps, trigger hooks, or perform actions that need to be observable, auditable, and reliable.

These are complementary, not competing, approaches.

🔍 Two Ways to Build Agentic AI

1. The “Chat + Tools” Approach (OpenAI, Anthropic, MCP)

The LLM drives everything.

response = client.chat.completions.create(
  model="gpt-4o",
  messages=[{"role": "user", "content": "What’s the weather in Berlin?"}],
  tools=[weather_tool]
)
Enter fullscreen mode Exit fullscreen mode
  • The model decides whether to call a tool.
  • Your code runs it and returns the result.
  • The model gives a final answer.

✅ Great for

  • Quick Q&A
  • Simple assistants
  • Early prototypes

❌ Falls short when you need

  • Multi-step logic
  • Retries or human approval
  • Audit trails or state
  • Compliance or safety guardrails

Here, the LLM is both brain and driver. You hand it tools and hope for the best. (We’ve all seen what happens when an unguarded LLM calls a destructive tool.)

2. The “Workflow Runtime” Approach (contenox, Temporal+LLMs, custom orchestrators)

I’ll use contenox to show how this works differently.

You define the workflow as a clear sequence of tasks. Each has a handler, optional LLM use, and transitions.

Realistic contenox syntax:

id: weather-advisor
description: Suggests actions based on the weather forecast
tasks:
  - id: get_weather
    description: Fetch weather data via external hook
    handler: hook
    hook:
      name: weather-api
      tool_name: get_forecast
      args:
        location: "Berlin"
    output_template: "{{.temperature}}"
    transition:
      branches:
        - operator: ">"
          when: "25"
          goto: "suggest_icecream"
        - operator: "default"
          goto: "suggest_walk"

  - id: suggest_icecream
    handler: model_execution
    system_instruction: "If it's hot, suggest a fun outdoor activity involving ice cream."
    execute_config:
      model: phi3:3.8b
      provider: ollama
    transition:
      branches:
        - operator: "default"
          goto: "end"

  - id: suggest_walk
    handler: model_execution
    system_instruction: "If it's cool, suggest something relaxing like a walk or a coffee indoors."
    execute_config:
      model: phi3:3.8b
      provider: ollama
    transition:
      branches:
        - operator: "default"
          goto: "end"
Enter fullscreen mode Exit fullscreen mode

✅ Great for

  • Reliable, stateful workflows
  • Real actions (APIs, notifications, DB writes)
  • Replay, audit, and debugging
  • Controlled, compliant agents

❌ Overkill for

  • Simple chatbots
  • One-off prompts

Here, you control the flow. The LLM is just one worker in the chain.

🧠 Why Both Exist

They solve different problems.

Context Goal Best Tool
Assistive AI “Help me get an answer fast.” OpenAI + Tools / MCP
Autonomous AI “Run a safe, reliable process.” Workflow runtime (contenox, Flyte, Temporal)

Think of it this way:

  • OpenAI + Tools is your clever intern—fast but unpredictable.
  • contenox is your project manager—structured, logged, and accountable.

🛠️ How to Choose

Use Case Best Approach
“Ask HR about PTO policy.” ✅ OpenAI + RAG
“Detect outage → Slack alert → Jira ticket → confirm fix.” ✅ contenox
“Generate a report and email it.” ⚠️ Start with OpenAI. Switch to contenox if reliability matters.
“Run AI in an air-gapped system.” ✅ contenox
“Weekend agent hack.” ✅ OpenAI + function calling

🔮 The Future: They’ll Meet in the Middle

MCP will add light state.

Workflow runtimes will simplify small jobs.

But the core question stays the same:

Is your AI assisting—or acting?

If it’s assisting, use tools.

If it’s acting, use orchestration.

🚀 Try contenox Yourself

It’s open source and self-hostable.

git clone https://github.com/contenox/runtime.git  
cd runtime
docker compose up -d
./scripts/bootstrap.sh nomic-embed-text:latest phi3:3.8b phi3:3.8b
Enter fullscreen mode Exit fullscreen mode

Define a workflow like the YAML above. Register your hooks. Watch your AI take real, safe actions.

🔗 GitHub: contenox/runtime

Top comments (0)