Agent loop

The core cycle an agent runs: observe context, choose an action (often a tool call), apply the result, and repeat until the task is done or a stop condition hits.

Learn

When to use it

Use an agent loop when a single model call isn't sufficient to complete a task. Agent loops enable the agent to observe, decide, act, and iterate, allowing for complex workflows like "monitor a system, trigger alerts, and adjust configurations until stability is achieved."

Quick example

In ChatGPT, when automating customer support, the model needs to assess user queries, decide on the best response or action, and continue the conversation until the issue is resolved. The agent loop in ChatGPT executes this cycle, enabling the model to handle ongoing interactions dynamically. This loop is what allows ChatGPT to maintain context and respond appropriately over multiple turns.

query → agent loop → response → repeat until issue resolved → stop

Ecosystem

Agent loops interact with tools and memory systems to maintain context and execute tasks effectively.

        ┌─ tools ↺─┐
context →│ agent loop │→ stop
        └─ memory ──┘

Misconceptions

Misconception	Rebuttal
It's a one-time process	It's a continuous cycle until a stop condition is met
Only for simple tasks	Handles complex, multi-step workflows

Trade-offs

Adaptability — requires careful management of context and state
Task completion — can lead to token burn if not efficiently managed