Differences between AI workflows and autonomous agents

AI workflows and autonomous AI agents both use large language models (LLM) but we shall looks at the differences here:

Comparisons

AI workflows

AI workflows use LLMs that can self-learn and do a single task autonomously, but the workflows (the sets of tasks) would still consist of:

• a pre-determined arrangement of steps
• a human-guided process
• a human-determined "definition of done"

examples include:

• a ChatGPT session summarizing simple documents (RAG)
• a coding chatbot that requires a developer's prompt

Autonomous AI agents

Meanwhile, autonomous AI agents consist of:

• autonomous ways of problem solving
• self-determined processes via external sources
• self-guided evaluation of processes and "definition of done"

examples include:

• coding agents that can determine parts of code that require optimization (e.g. website navigation or "buy now" button locations) based on changes in data flows and actual user habits
• making travel plans and booking travel services from various travel vendors
• autonomous research assistants browsing the internet and summarizing discoveries across multiple fields of study

Beyond its own LLM, an autonomous AI agent:

• will not only use its power to get a solution
  • but use its power to find its own way to get the solution!

Considerations

For simpler tasks:

• straightforward prompts to an LLM with (optionally) an AI workflow could do the trick
  • we do not need to automate everything!

For less consistent and more complex tasks:

• consider using agentic AI
  • while remaining mindful of the latency and costs that arise with agents

Summary

The main differences between "workflow" and "agent":

• workflows have less autonomy on the set of tasks than an agent
  • both use an autonomous LLM to perform a task
• the nature of an agent falls more on a spectrum than a hard line
  • an agent tends to have a higher degree of autonomy
• agents have the downside of increased latency and higher costs