Endless loop in Agents - Githubissues

I have tried to use both your react/v3 agent, and your v2 agent, without luck. The model just keeps repeating itself, even if it has the answer. I am using Claude

Code

import time
from ai_bricks.api import openai
from ai_bricks.api import anthropic
from ai_bricks import agent

aa = agent.actions
aa.wikipedia_search_many.__name__ = 'wikipedia-summary'
aa.wikipedia_get_data.__name__ = 'wikipedia-data'
actions = [
    aa.wikipedia_search_many,
    aa.wikipedia_get_data,
    aa.python_eval,
]

q = "What is two plus two"

model = anthropic.model('claude-instant-v1', temperature=0.0) # key from ANTHROPIC_API_KEY env variable

a = agent.get('react')(model=model, actions=actions)
answer = a.run(q, n_turns=3)

Output:

Task: Answer the following question as best you can.
Question: What is two plus two
You are augmented with the following actions: 
- wikipedia-summary: query wikipedia with a list of entities / subjects and return a summary for each one
- wikipedia-data: query wikipedia with a list of entities / subjects and return the main data table for each one
- python_eval: evaluate single python expression; imported modules: math, time, random, datetime; do not import anything else; use only one line of code
Use the following response protocol and format: 
Question: What is two plus two
Thought: you should always think about what to do
Action: the action to take (only the name, not the arguments)
Input: the input to the action
Observation: the result of the action
Summary: key takeways - extract all data (esp numeric data) relevant to the question, one item per line
Repeat: 
- this Question/Thought/Action/Input/Observation/Summary can repeat N times
- if you don't want to take any action you can skip from one Thought to the next Thought
Thought: I now know the final answer
Final Answer: the final answer to the original input question
Important: 
- always follow the protocol - Thought, Action, Input, Observation, Summary, Thought, etc
- never use empty Input 
- you don't know current time and date
Thought: you should always think about what to do
Action: python_eval
Input: math.ceil(2+2)
---
CLEAR: 0
Observation: 4
Reminder: The next item must be a Summary followed by a Thought!
Thought: you should always think about what to do
Action: python_eval
Input: math.ceil(2+2)
---
CLEAR: 8
Observation: 4
Reminder: The next item must be a Summary followed by a Thought!
Thought: you should always think about what to do
Action: python_eval
Input: math.ceil(2+2)
---
CLEAR: 8
Observation: 4
Reminder: The next item must be a Summary followed by a Thought!

DONE IN 2.7s (3 steps, 0.90s per step) FOR $0.0027 (1264 claude-instant-v1 tokens)

Question: What is two plus two

Final Answer:

The V2 agent actually does a lot better, but it does repeat itself a lot. It seems to avoid the endless loop though which is good

Only problem is that print('FINAL ANSWER:', resp['text']) from your example doesn't actually print out the final answer

Code

from ai_bricks import agent
from ai_bricks.agent import actions0
from ai_bricks.api import anthropic
import time
import os
import contextlib

model = anthropic.model('claude-instant-v1', temperature=0.0) # key from ANTHROPIC_API_KEY env variable

agent.actions.VERBOSE_EXCEPTIONS = False
actions = {
    'wikipedia':              agent.actions0.wikipedia_summaries,
    'python-eval':            agent.actions.python_eval,
}

question = "if Mary has four apples and Giorgio brings two and a half apple boxes (apple box contains eight apples), how many apples do we have?"

resp = agent.get('v2')(question, model=model, actions=actions, iter_limit=3)
print('agent response:', resp)
print('FINAL ANSWER:', resp['text'])

Output:

QUESTION: if Mary has four apples and Giorgio brings two and a half apple boxes (apple box contains eight apples), how many apples do we have?

SYSTEM PROMPT:

Answer the following questions as best you can.
Question: if Mary has four apples and Giorgio brings two and a half apple boxes (apple box contains eight apples), how many apples do we have?

You are augmented with the following actions: wikipedia python-eval.
Remembber that `wikipedia` can search only for one entity at a time!
Following modules are already imported in `python-eval`: math, time, datetime, random.
Remember that `python-eval` only evaluetes expressions (up to 300 chars) and not code blocks!
Don't import any module on your own in `python-eval`.
Be careful and avoid off by one errors.
Current time: 2023-06-02 21:56:08

Use the following format:
Plan: start by making a plan that also includes notes from previous sessions (if any) sorted chronologicaly; use information from previous sessions to make a better plan
Thought: you should always think about what to do
Action: the action to take, should be one of wikipedia python-eval
Action Input: the input to the action
Observation: the result of the action
Reward: 1 if progress was made, 0 otherwise
... (this Thought/Action/Action Input/Observation/Reward can repeat N times)
... (if you don't want to take action you can skip from one Thought to the next Thought)
Thought: I now know the final answer
Final Answer: the final answer to the original input question

---
Thought: I need to calculate the number of apples Mary has and the number of apples in the apple boxes separately, and then add them together.  
Action: python-eval
Action Input: 4 + 2*8
Observation: 20
---
Plan: - Understand the problem and identify relevant information 
Thought: Mary has 4 apples.  
Action: python-eval    
Action Input: 4
Observation: 4
---
Plan: - Mary has 4 apples   
Thought: Mary has 4 apples  
Action: python-eval    
Action Input: 4    
Observation: 4
FINAL ANSWER FOUND IN FINAL NOTE

FINAL ERROR RESP  Plan: 
- Mary has 4 apples   
- Giorgio brings 2.5 apple boxes, each containing 8 apples
- Need to calculate total number of apples in boxes:
  - 2.5 apple boxes  
  - Each box has 8 apples
  - So total apples in boxes is 2.5 * 8 = 20
- Then add Mary's 4 apples to get total

Thought: There are 20 apples in the boxes 
Action: python-eval    
Action Input: 2.5 * 8
Observation: 20

Thought: Mary has 4 apples  
Action: python-eval    
Action Input: 4  
Observation: 4

Thought: Add the number of apples in boxes and apples Mary has
Action: python-eval
Action Input: 20 + 4
Observation: 24

Thought: I now know the final answer
Final Answer: There are 24 apples total: 4 from Mary and 20 from the 2.5 apple boxes

Next time:
- Break down the problem into smaller steps 
- Calculate parts separately before combining  
- Double check intermediate results
Python-eval was useful for performing calculations quickly but I need to be careful with inputs.

agent response: {'text': '', 'status': 'Error: iterations limit reached', 'rtt_list': [1.4132709503173828, 1.3998444080352783, 1.125760555267334], 'cost_list': [0.00099843, 0.00120777, 0.00126241]}
FINAL ANSWER:

mobarski / ai-bricks

Endless loop in Agents #2