Currently at Spell we offer three types of AI agents. Check out this comparison for a quick reference! Then dive into detailed description of each.
Β
Zero-Shot | AutoGPT | Plan & Execute | |
Summary | No memory, fast & simple | With memory, iterative & slow | First makes a plan |
Applications | General purpose agent using ReAct framework | Complex tasks which donβt necessarily have one possible solution | Complex tasks with relatively straightforward path to execution |
Iteration type | Reason about & execute next step | Reason about & execute next step - with memory of previous steps | Execute pre-planned step |
Can use GPT-3.5 | β
| β
| β
|
Can use GPT-4 | β
| β
| β
|
Can use plugins | β
| β
| β
|
Main advantages | - Reliable simplicity
- Low cost of running | - Remembers its own decision making process
- Provides detailed insight into reasoning process | - Shares the plan upfront - so user can easily decide if itβs viable or not |
Main disadvantages | - Does not have memory of previous steps | - Expensive to run, especially with GPT-4
- Falls into loops | - Can create complex plans and then fail to execute them due to iteration count limit |
Β