4 Design Dimensions of Agentic Workflows

People often mix up AI agents and AI workflows. Some say any LLM with tools is an agent. Others say an agent must run alone for a long time. These debates do not help you build better systems.

Stop asking if a system is agentic. Instead, look at these four design dimensions:

  1. Who decides the next step?
  • Code-driven: Code controls the sequence. Use this for high-stakes tasks with clear rules.
  • Model-driven: The LLM plans the steps. Use this for open-ended tasks.
  • Hybrid: Code sets the frame and the LLM makes small decisions. This is the most practical choice.
  1. Is the path fixed?
  • Fixed: Steps follow a set order.
  • Conditional: The system uses branches or retries based on results.
  • Adaptive: The path changes at runtime. Use this for research or debugging.
  1. How do agents work together?
  • Single Agent: One agent uses multiple tools. This is simple but context can get messy.
  • Manager-Worker: One agent breaks down tasks and gives them to others.
  • Handoff: One agent passes control to a specialist.
  • Peer: Multiple agents collaborate without a central leader. This is the hardest to debug.
  1. Where do humans step in?
  • Human-triggered: You approve every major action.
  • Checkpointed: The system works alone but stops for approval at high-risk steps.
  • Goal-driven with guardrails: You set the goal and the system works within limits like budgets and permissions.

Do not add agents just to be fancy. Adding an agent increases latency, cost, and debugging difficulty. Use multiple agents only if you can run tasks in parallel or if a single agent has too much context.

Match your architecture to the task. If the task is risky, use checkpoints. If the task is simple, use code.

Ask yourself these six questions before building:

  • Can regular code solve this?
  • How much variation is in the input?
  • Can the task be split up?
  • How much does a failure cost?
  • Can you see where it breaks?
  • Is the extra complexity worth the result?

A good workflow stays flexible where it must and stays predictable where it needs to.

Source: https://dev.to/whchi/4-design-dimensions-of-agentic-workflows-1i0m

Optional learning community: https://t.me/GyaanSetuAi