Back to full roadmap
topicadvanced
Agentic Loop Architecture
while(!done) { think → act → observe → update_state } — backbone of modern agents.
5 hours1 resources1 prereqs
An agent ≠ single LLM call. There's a loop:
- Plan — break goal into steps
- Select tool — pick the right one
- Execute — run it
- Observe — interpret result
- Update — revise the plan
- Goal done? No → back to 1
Dangers:
- Infinite loop → set max iterations
- Complex tool chains → trace logging is mandatory
- Bad tool output → retry + fallback