# 4 Types of AI Agent Loops: From Human-Led to Autonomous

## Executive summary

This video details four types of AI agent loops—interactive, delegated, scheduled, and event-driven—which define the level of autonomy in agentic workflows. All loops follow a basic cycle (Goal, Action, Observe, Adjust), but they differ in who sets the goal and who determines the stopping condition, ranging from human-led tasks to fully autonomous systems.

## Key takeaways

- Interactive Loop (Human-Led): The human sets the goal and judges the outcome at the end of every cycle, making it suitable for fuzzy tasks lacking a measurable finish line. This loop is difficult to scale for repetitive jobs.
- Delegated Loop (AI-Judged): The human sets the goal with a specific stopping condition, which is judged by an AI check. The agent runs until the condition is met, after which the result is provided for human review.
- Scheduled Loop (Time-Driven): The loop is initiated by a clock, not a person. It uses an AI check to judge the stopping condition, making it ideal for recurrent or time-sensitive tasks, such as managing out-of-date libraries.
- Event-Driven Loop (Autonomous): The loop is triggered by an external event (e.g., a pull request). It uses a second, dedicated reviewer agent to check the stopping condition, with human review reserved only for exceptions.

## Technical details

- Agentic Workflow Cycle: All agentic loops share the pattern: Agent takes a Goal $ ightarrow$ Performs an Action (e.g., running a command) $ ightarrow$ Observes the output $ ightarrow$ Adjusts based on observation, repeating until completion.
- Stopping Conditions: The stopping condition determines when the loop exits. In delegated and scheduled loops, this check is performed by an AI. In event-driven loops, a separate reviewer agent performs the check.
- Human-in-the-Loop vs. AI Review: The interactive loop requires human judgment at the end of the process. The event-driven loop replaces the final human review with an AI reviewer agent, reserving human intervention only for exceptions.

## Practical implications

- When automating build processes (e.g., dependency upgrades), a build engineer can select the appropriate loop type: use Interactive for ambiguous, exploratory tasks, Delegated/Scheduled for routine maintenance, and Event-Driven for reactive CI/CD triggers (like PR creation).
- By delegating the stopping condition check to an AI, the system can achieve greater scalability compared to requiring human judgment after every iteration.

## Topics

AI Agents, Agentic AI, CI/CD, Automation, Workflow Engineering, Loop Engineering, IBM Technology

Source: https://www.youtube.com/watch?v=4LINhsEZEDI
