From Knowledge to Tales: Code Brokers for KPI Narratives

, we frequently want to research what’s happening with KPIs: whether or not we’re reacting to anomalies on our dashboards or simply routinely doing a numbers replace. Based mostly on my years of expertise as a KPI analyst, I might estimate that greater than 80% of those duties are pretty customary and could be solved simply by following a easy guidelines. 

Right here’s a high-level plan for investigating a KPI change (you could find extra particulars within the article “Anomaly Root Trigger Evaluation 101”):

• Estimate the top-line change within the metric to know the magnitude of the shift. 
• Test information high quality to make sure that the numbers are correct and dependable.
• Collect context about inside and exterior occasions which may have influenced the change.
• Slice and cube the metric to establish which segments are contributing to the metric’s shift.
• Consolidate your findings in an government abstract that features hypotheses and estimates of their impacts on the primary KPI.

Since we now have a transparent plan to execute, such duties can doubtlessly be automated utilizing AI brokers. The code brokers we not too long ago mentioned could possibly be a very good match there, as their means to write down and execute code will assist them to analyse information effectively, with minimal back-and-forth. So, let’s attempt constructing such an agent utilizing the HuggingFace smolagents framework. 

Whereas engaged on our job, we’ll talk about extra superior options of the smolagents framework:

• Strategies for tweaking every kind of prompts to make sure the specified behaviour.
• Constructing a multi-agent system that may clarify the Kpi modifications and hyperlink them to root causes. 
• Including reflection to the circulation with supplementary planning steps.

MVP for explaining KPI modifications

As ordinary, we’ll take an iterative strategy and begin with a easy MVP, specializing in the slicing and dicing step of the evaluation. We are going to analyse the modifications of a easy metric (income) break up by one dimension (nation). We are going to use the dataset from my earlier article, “Making sense of KPI modifications”.

Let’s load the information first. 

raw_df = pd.read_csv('absolute_metrics_example.csv', sep = 't')
df = raw_df.groupby('nation')[['revenue_before', 'revenue_after_scenario_2']].sum()
  .sort_values('revenue_before', ascending = False).rename(
    columns = {'revenue_after_scenario_2': 'after', 
      'revenue_before': 'earlier than'})

Subsequent, let’s initialise the mannequin. I’ve chosen the OpenAI GPT-4o-mini as my most well-liked possibility for easy duties. Nonetheless, the smolagents framework helps every kind of fashions, so you should utilize the mannequin you like. Then, we simply must create an agent and provides it the duty and the dataset.

from smolagents import CodeAgent, LiteLLMModel

mannequin = LiteLLMModel(model_id="openai/gpt-4o-mini", 
  api_key=config['OPENAI_API_KEY']) 

agent = CodeAgent(
    mannequin=mannequin, instruments=[], max_steps=10,
    additional_authorized_imports=["pandas", "numpy", "matplotlib.*", 
      "plotly.*"], verbosity_level=1 
)

job = """
Here's a dataframe exhibiting income by section, evaluating values 
earlier than and after.
Might you please assist me perceive the modifications? Particularly:
1. Estimate how the overall income and the income for every section 
have modified, each in absolute phrases and as a proportion.
2. Calculate the contribution of every section to the overall 
change in income.

Please spherical all floating-point numbers within the output 
to 2 decimal locations.
"""

agent.run(
    job,
    additional_args={"information": df},
)

The agent returned fairly a believable outcome. We received detailed statistics on the metric modifications in every section and their influence on the top-line KPI.

{'total_before': 1731985.21, 'total_after': 
1599065.55, 'total_change': -132919.66, 'segment_changes': 
{'absolute_change': {'different': 4233.09, 'UK': -4376.25, 'France': 
-132847.57, 'Germany': -690.99, 'Italy': 979.15, 'Spain': 
-217.09}, 'percentage_change': {'different': 0.67, 'UK': -0.91, 
'France': -55.19, 'Germany': -0.43, 'Italy': 0.81, 'Spain': 
-0.23}, 'contribution_to_change': {'different': -3.18, 'UK': 3.29, 
'France': 99.95, 'Germany': 0.52, 'Italy': -0.74, 'Spain': 0.16}}}

Let’s check out the code generated by the agent. It’s tremendous, however there’s one potential challenge. The Llm recreated the dataframe primarily based on the enter information as an alternative of referencing it straight. This strategy shouldn’t be best (particularly when working with large datasets), as it may result in errors and better token utilization.

import pandas as pd                                                                                                        
 
# Creating the DataFrame from the offered information                 
information = {                                                        
    'earlier than': [632767.39, 481409.27, 240704.63, 160469.75,      
120352.31, 96281.86],                                           
    'after': [637000.48, 477033.02, 107857.06, 159778.76,       
121331.46, 96064.77]                                            
}                                                               
index = ['other', 'UK', 'France', 'Germany', 'Italy', 'Spain']  
df = pd.DataFrame(information, index=index)                            
                                                                
# Calculating whole income earlier than and after                    
total_before = df['before'].sum()                               
total_after = df['after'].sum()                                 
                                                                
# Calculating absolute and proportion change for every section   
df['absolute_change'] = df['after'] - df['before']              
df['percentage_change'] = (df['absolute_change'] /              
df['before']) * 100                                             
                                                                
# Calculating whole income change                              
total_change = total_after - total_before                       
                                                                
# Calculating contribution of every section to the overall change  
df['contribution_to_change'] = (df['absolute_change'] /         
total_change) * 100                                             
                                                                
# Rounding outcomes                                              
df = df.spherical(2)                                                
                                                                
# Printing the calculated outcomes                               
print("Complete income earlier than:", total_before)                    
print("Complete income after:", total_after)                      
print("Complete change in income:", total_change)                 
print(df)

It’s price fixing this downside earlier than shifting on to constructing a extra complicated system.

Tweaking prompts

Because the LLM is simply following the directions given to it, we’ll tackle this challenge by tweaking the immediate.

Initially, I tried to make the duty immediate extra express, clearly instructing the LLM to make use of the offered variable.

job = """Here's a dataframe exhibiting income by section, evaluating 
values earlier than and after. The info is saved in df variable. 
Please, use it and do not attempt to parse the information your self. 

Might you please assist me perceive the modifications?
Particularly:
1. Estimate how the overall income and the income for every section 
have modified, each in absolute phrases and as a proportion.
2. Calculate the contribution of every section to the overall change in income.

Please spherical all floating-point numbers within the output to 2 decimal locations.
"""

It didn’t work. So, the subsequent step is to look at the system immediate and see why it really works this fashion. 

print(agent.prompt_templates['system_prompt'])

#... 
# Listed below are the foundations you must at all times comply with to unravel your job:
# 1. At all times present a 'Thought:' sequence, and a 'Code:n```py' sequence ending with '```' sequence, else you'll fail.
# 2. Use solely variables that you've got outlined.
# 3. At all times use the precise arguments for the instruments. DO NOT move the arguments as a dict as in 'reply = wiki({'question': "What's the place the place James Bond lives?"})', however use the arguments straight as in 'reply = wiki(question="What's the place the place James Bond lives?")'.
# 4. Take care to not chain too many sequential instrument calls in the identical code block, particularly when the output format is unpredictable. For example, a name to go looking has an unpredictable return format, so should not have one other instrument name that is dependent upon its output in the identical block: relatively output outcomes with print() to make use of them within the subsequent block.
# 5. Name a instrument solely when wanted, and by no means re-do a instrument name that you simply beforehand did with the very same parameters.
# 6. Do not identify any new variable with the identical identify as a instrument: as an example do not identify a variable 'final_answer'.
# 7. By no means create any notional variables in our code, as having these in your logs will derail you from the true variables.
# 8. You should utilize imports in your code, however solely from the next record of modules: ['collections', 'datetime', 'itertools', 'math', 'numpy', 'pandas', 'queue', 'random', 're', 'stat', 'statistics', 'time', 'unicodedata']
# 9. The state persists between code executions: so if in a single step you have created variables or imported modules, these will all persist.
# 10. Do not surrender! You are answerable for fixing the duty, not offering instructions to unravel it.
# Now Start!

On the finish of the immediate, we now have the instruction "# 2. Use solely variables that you've got outlined!". This is likely to be interpreted as a strict rule to not use another variables. So, I modified it to "# 2. Use solely variables that you've got outlined or ones offered in further arguments! By no means attempt to copy and parse further arguments." 

modified_system_prompt = agent.prompt_templates['system_prompt']
    .substitute(
        '2. Use solely variables that you've got outlined!', 
        '2. Use solely variables that you've got outlined or ones offered in further arguments! By no means attempt to copy and parse further arguments.'
    )
agent.prompt_templates['system_prompt'] = modified_system_prompt

This alteration alone didn’t assist both. Then, I examined the duty message. 

╭─────────────────────────── New run ────────────────────────────╮
│                                                                │
│ Here's a pandas dataframe exhibiting income by section,         │
│ evaluating values earlier than and after.                             │
│ Might you please assist me perceive the modifications?               │
│ Particularly:                                                  │
│ 1. Estimate how the overall income and the income for every     │
│ section have modified, each in absolute phrases and as a          │
│ proportion.                                                    │
│ 2. Calculate the contribution of every section to the overall     │
│ change in income.                                             │
│                                                                │
│ Please spherical all floating-point numbers within the output to 2   │
│ decimal locations.                                                │
│                                                                │
│ You will have been supplied with these further arguments, that   │
│ you may entry utilizing the keys as variables in your python      │
│ code:                                                          │
│ {'df':             earlier than      after                           │
│ nation                                                        │
│ different    632767.39  637000.48                                  │
│ UK       481409.27  477033.02                                  │
│ France   240704.63  107857.06                                  │
│ Germany  160469.75  159778.76                                  │
│ Italy    120352.31  121331.46                                  │
│ Spain     96281.86   96064.77}.                                │
│                                                                │
╰─ LiteLLMModel - openai/gpt-4o-mini ────────────────────────────╯

It has an instruction associated the the utilization of further arguments "You will have been supplied with these further arguments, that you could entry utilizing the keys as variables in your python code". We will attempt to make it extra particular and clear. Sadly, this parameter shouldn’t be uncovered externally, so I needed to find it in the supply code. To search out the trail of a Python package deal, we will use the next code.

import smolagents 
print(smolagents.__path__)

Then, I discovered the brokers.py file and modified this line to incorporate a extra particular instruction.

self.job += f"""
You will have been supplied with these further arguments accessible as variables 
with names {",".be a part of(additional_args.keys())}. You possibly can entry them straight. 
Here's what they include (only for informational functions): 
{str(additional_args)}."""

It was a little bit of hacking, however that’s typically what occurs with the LLM frameworks. Don’t overlook to reload the package deal afterwards, and we’re good to go. Let’s take a look at whether or not it really works now.

job = """
Here's a pandas dataframe exhibiting income by section, evaluating values 
earlier than and after. 

Your job shall be perceive the modifications to the income (after vs earlier than) 
in numerous segments and supply government abstract.
Please, comply with the next steps:
1. Estimate how the overall income and the income for every section 
have modified, each in absolute phrases and as a proportion.
2. Calculate the contribution of every section to the overall change 
in income.

Spherical all floating-point numbers within the output to 2 decimal locations. 
"""
agent.logger.stage = 1 # Decrease verbosity stage
agent.run(
    job,
    additional_args={"df": df},
)

Hooray! The issue has been mounted. The agent now not copies the enter variables and references df variable straight as an alternative. Right here’s the newly generated code.

import pandas as pd                                             
                                                                  
# Calculate whole income earlier than and after                      
total_before = df['before'].sum()                               
total_after = df['after'].sum()                                 
total_change = total_after - total_before                       
percentage_change_total = (total_change / total_before * 100)   
if total_before != 0 else 0                                     
                                                                
# Spherical values                                                  
total_before = spherical(total_before, 2)                           
total_after = spherical(total_after, 2)                             
total_change = spherical(total_change, 2)                           
percentage_change_total = spherical(percentage_change_total, 2)     
                                                                
# Show outcomes                                               
print(f"Complete Income Earlier than: {total_before}")                  
print(f"Complete Income After: {total_after}")                    
print(f"Complete Change: {total_change}")                          
print(f"Share Change: {percentage_change_total}%")

Now, we’re prepared to maneuver on to constructing the precise agent that can remedy our job.

AI agent for KPI narratives

Lastly, it’s time to work on the AI agent that can assist us clarify KPI modifications and create an government abstract.

Our agent will comply with this plan for the foundation trigger evaluation:

• Estimate the top-line KPI change. 
• Slice and cube the metric to know which segments are driving the shift. 
• Search for occasions within the change log to see whether or not they can clarify the metric modifications.
• Consolidate all of the findings within the complete government abstract.

After loads of experimentation and a number of other tweaks, I’ve arrived at a promising outcome. Listed below are the important thing changes I made (we’ll talk about them intimately later):

• I leveraged the multi-agent setup by including one other workforce member — the change log Agent, who can entry the change log and help in explaining KPI modifications.
• I experimented with extra highly effective fashions like gpt-4o and gpt-4.1-mini since gpt-4o-mini wasn’t enough. Utilizing stronger fashions not solely improved the outcomes, but additionally considerably diminished the variety of steps: with gpt-4.1-miniI received the ultimate outcome after simply six steps, in comparison with 14–16 steps with gpt-4o-mini. This means that investing in dearer fashions is likely to be worthwhile for agentic workflows.
• I offered the agent with the complicated instrument to analyse KPI modifications for easy metrics. The instrument performs all of the calculations, whereas LLM can simply interpret the outcomes. I mentioned the strategy to KPI modifications evaluation intimately in my earlier article. 
• I reformulated the immediate into a really clear step-by-step information to assist the agent keep on observe. 
• I added planning steps that encourage the LLM agent to suppose via its strategy first and revisit the plan each three iterations. 

After all of the changes, I received the next abstract from the agent, which is fairly good.

Government Abstract:
Between April 2025 and Could 2025, whole income declined sharply by
roughly 36.03%, falling from 1,731,985.21 to 1,107,924.43, a
drop of -624,060.78 in absolute phrases.
This decline was primarily pushed by vital income 
reductions within the 'new' buyer segments throughout a number of 
nations, with declines of roughly 70% in these segments.

Probably the most impacted segments embrace:
- other_new: earlier than=233,958.42, after=72,666.89, 
abs_change=-161,291.53, rel_change=-68.94%, share_before=13.51%, 
influence=25.85, impact_norm=1.91
- UK_new: earlier than=128,324.22, after=34,838.87, 
abs_change=-93,485.35, rel_change=-72.85%, share_before=7.41%, 
influence=14.98, impact_norm=2.02
- France_new: earlier than=57,901.91, after=17,443.06, 
abs_change=-40,458.85, rel_change=-69.87%, share_before=3.34%, 
influence=6.48, impact_norm=1.94
- Germany_new: earlier than=48,105.83, after=13,678.94, 
abs_change=-34,426.89, rel_change=-71.56%, share_before=2.78%, 
influence=5.52, impact_norm=1.99
- Italy_new: earlier than=36,941.57, after=11,615.29, 
abs_change=-25,326.28, rel_change=-68.56%, share_before=2.13%, 
influence=4.06, impact_norm=1.91
- Spain_new: earlier than=32,394.10, after=7,758.90, 
abs_change=-24,635.20, rel_change=-76.05%, share_before=1.87%, 
influence=3.95, impact_norm=2.11

Based mostly on evaluation from the change log, the primary causes for this 
pattern are:
1. The introduction of latest onboarding controls carried out on Could 
8, 2025, which diminished new buyer acquisition by about 70% to 
forestall fraud.
2. A postal service strike within the UK beginning April 5, 2025, 
inflicting order supply delays and elevated cancellations 
impacting the UK new section.
3. A rise in VAT by 2% in Spain as of April 22, 2025, 
affecting new buyer pricing and inflicting greater cart 
abandonment.

These elements mixed clarify the outsized unfavourable impacts 
noticed in new buyer segments and the general income decline.

The LLM agent additionally generated a bunch of illustrative charts (they have been a part of our development explaining instrument). For instance, this one reveals the impacts throughout the mixture of nation and maturity.

The outcomes look actually thrilling. Now let’s dive deeper into the precise implementation to know the way it works below the hood. 

Multi-AI agent setup

We are going to begin with our change log agent. This agent will question the change log and attempt to establish potential root causes for the metric modifications we observe. Since this agent doesn’t must do complicated operations, we implement it as a ToolCallingAgent. As a result of this agent shall be known as by one other agent, we have to outline its identify and description attributes.

@instrument 
def get_change_log(month: str) -> str: 
    """
    Returns the change log (record of inside and exterior occasions which may have affected our KPIs) for the given month 

    Args:
        month: month within the format %Y-%m-01, for instance, 2025-04-01
    """
    return events_df[events_df.month == month].drop('month', axis = 1).to_dict('data')

mannequin = LiteLLMModel(model_id="openai/gpt-4.1-mini", api_key=config['OPENAI_API_KEY'])
change_log_agent = ToolCallingAgent(
    instruments=[get_change_log],
    mannequin=mannequin,
    max_steps=10,
    identify="change_log_agent",
    description="Helps you discover the related info within the change log that may clarify modifications on metrics. Present the agent with all of the context to obtain data",
)

Because the supervisor agent shall be calling this agent, we gained’t have any management over the question it receives. Due to this fact, I made a decision to change the system immediate to incorporate further context.

change_log_system_prompt = '''
You are a grasp of the change log and also you assist others to elucidate 
the modifications to metrics. Whenever you obtain a request, lookup the record of occasions 
occurred by month, then filter the related info primarily based 
on offered context and return again. Prioritise probably the most possible elements 
affecting the KPI and restrict your reply solely to them.
'''

modified_system_prompt = change_log_agent.prompt_templates['system_prompt'] 
  + 'nnn' + change_log_system_prompt

change_log_agent.prompt_templates['system_prompt'] = modified_system_prompt

To allow the first agent to delegate duties to the change log agent, we merely must specify it within the managed_agents area.

agent = CodeAgent(
    mannequin=mannequin,
    instruments=[calculate_simple_growth_metrics],
    max_steps=20,
    additional_authorized_imports=["pandas", "numpy", "matplotlib.*", "plotly.*"],
    verbosity_level = 2, 
    planning_interval = 3,
    managed_agents = [change_log_agent]
)

Let’s see the way it works. First, we will take a look at the brand new system immediate for the first agent. It now contains details about workforce members and directions on ask them for assist.

You too can give duties to workforce members.
Calling a workforce member works the identical as for calling a instrument: merely, 
the one argument you may give within the name is 'job'.
Provided that this workforce member is an actual human, you need to be very verbose 
in your job, it must be a protracted string offering informations 
as detailed as vital.
Here's a record of the workforce members that you could name:
```python
def change_log_agent("Your question goes right here.") -> str:
    """Helps you discover the related info within the change log that 
    can clarify modifications on metrics. Present the agent with all of the context 
    to obtain data"""
```

The execution log reveals that the first agent efficiently delegated the duty to the second agent and acquired the next response.

<-- Main agent calling the change log agent -->

─ Executing parsed code: ─────────────────────────────────────── 
  # Question change_log_agent with the detailed job description     
  ready                                                        
  context_for_change_log = (                                      
      "We analyzed modifications in income from April 2025 to Could      
  2025. We discovered giant decreases "                                
      "primarily within the 'new' maturity segments throughout nations:    
  Spain_new, UK_new, Germany_new, France_new, Italy_new, and      
  other_new. "                                                    
      "The income fell by round 70% in these segments, which    
  have outsized unfavourable influence on whole income change. "        
      "We need to know the 1-3 most possible causes for this      
  vital drop in income within the 'new' buyer segments      
  throughout this era."                                            
  )                                                               
                                                                  
  clarification = change_log_agent(job=context_for_change_log)     
  print("Change log agent clarification:")                          
  print(clarification)                                              
 ──────────────────────────────────────────────────────────────── 

<-- Change log agent execution begin -->
╭──────────────────── New run - change_log_agent ─────────────────────╮
│                                                                     │
│ You are a useful agent named 'change_log_agent'.                    │
│ You will have been submitted this job by your supervisor.                  │
│ ---                                                                 │
│ Activity:                                                               │
│ We analyzed modifications in income from April 2025 to Could 2025.         │
│ We discovered giant decreases primarily within the 'new' maturity segments      │
│ throughout nations: Spain_new, UK_new, Germany_new, France_new,       │
│ Italy_new, and other_new. The income fell by round 70% in these   │
│ segments, which have outsized unfavourable influence on whole income      │
│ change. We need to know the 1-3 most possible causes for this       │
│ vital drop in income within the 'new' buyer segments throughout   │
│ this era.                                                        │
│ ---                                                                 │
│ You are serving to your supervisor remedy a wider job: so be sure to     │
│ not present a one-line reply, however give as a lot info as      │
│ potential to present them a transparent understanding of the reply.          │
│                                                                     │
│ Your final_answer WILL HAVE to include these components:                 │
│ ### 1. Activity consequence (brief model):                                │
│ ### 2. Activity consequence (extraordinarily detailed model):                   │
│ ### 3. Extra context (if related):                            │
│                                                                     │
│ Put all these in your final_answer instrument, all the pieces that you simply do     │
│ not move as an argument to final_answer shall be misplaced.               │
│ And even when your job decision shouldn't be profitable, please return   │
│ as a lot context as potential, in order that your supervisor can act upon      │
│ this suggestions.                                                      │
│                                                                     │
╰─ LiteLLMModel - openai/gpt-4.1-mini ────────────────────────────────╯

Utilizing the smolagents framework, we will simply arrange a easy multi-agent system, the place a supervisor agent coordinates and delegates duties to workforce members with particular expertise. 

Iterating on the immediate

We’ve began with a really high-level immediate outlining the aim and a imprecise path, however sadly, it didn’t work persistently. LLMs will not be good sufficient but to determine the strategy on their very own. So, I created an in depth step-by-step immediate describing the entire plan and together with the detailed specs of the expansion narrative instrument we’re utilizing. 

job = """
Here's a pandas dataframe exhibiting the income by section, evaluating values 
earlier than (April 2025) and after (Could 2025). 

You are a senior and skilled information analyst. Your job shall be to know 
the modifications to the income (after vs earlier than) in numerous segments 
and supply government abstract.

## Comply with the plan:
1. Begin by udentifying the record of dimensions (columns in dataframe that 
will not be "earlier than" and "after")
2. There is likely to be a number of dimensions within the dataframe. Begin high-level 
by  every dimension in isolation, mix all outcomes 
collectively into the record of segments analysed (remember to save lots of 
the dimension used for every section). 
Use the offered instruments to analyse the modifications of metrics: {tools_description}. 
3. Analyse the outcomes from earlier step and preserve solely segments 
which have outsized influence on the KPI change (absolute of impact_norm 
is above 1.25). 
4. Test what dimensions are current within the record of serious section, 
if there are a number of ones - execute the instrument on their mixtures 
and add to the analysed segments. If after including a further dimension, 
all subsegments present shut different_rate and impact_norm values, 
then we will exclude this break up (despite the fact that impact_norm is above 1.25), 
because it does not clarify something. 
5. Summarise the numerous modifications you recognized. 
6. Attempt to clarify what's going on with metrics by getting data 
from the change_log_agent. Please, present the agent the complete context 
(what segments have outsized influence, what's the relative change and 
what's the interval we're ). 
Summarise the data from the changelog and point out 
solely 1-3 probably the most possible causes of the KPI change 
(ranging from probably the most impactful one).
7. Put collectively 3-5 sentences commentary what occurred high-level 
and why (primarily based on the information acquired from the change log). 
Then comply with it up with extra detailed abstract: 
- Prime-line whole worth of metric earlier than and after in human-readable format, 
absolute and relative change 
- Listing of segments that meaningfully influenced the metric positively 
or negatively with the next numbers: values earlier than and after, 
absoltue and relative change, share of section earlier than, influence 
and normed influence. Order the segments by absolute worth 
of absolute change because it represents the facility of influence. 

## Instruction on the calculate_simple_growth_metrics instrument:
By default, you must use the instrument for the entire dataset not the section, 
because it offers you the complete details about the modifications.

Right here is the steering  interpret the output of the instrument
- distinction - absolutely the distinction between after and earlier than values
- difference_rate - the relative distinction (if it is shut for 
  all segments then the dimension shouldn't be informative)
- influence - the share of KPI differnce defined by this section 
- segment_share_before - share of section earlier than
- impact_norm - influence normed on the share of segments, we're  
  in very excessive or very low numbers since they present outsized influence, 
  rule of thumb - impact_norm between -1.25 and 1.25 is not-informative 

In the event you're utilizing the instrument on the subset of dataframe take into account, 
that the outcomes will not be aplicable to the complete dataset, so keep away from utilizing it 
except you need to explicitly take a look at subset (i.e. change in France). 
In the event you determined to make use of the instrument on a selected section 
and share these ends in the manager abstract, explicitly define 
that we're diving deeper into a selected section.
""".format(tools_description = tools_description)
agent.run(
    job,
    additional_args={"df": df},
)

Explaining all the pieces in such element was fairly a frightening job, but it surely’s vital if we wish constant outcomes.

Planning steps

The smolagents framework permits you to add planning steps to your agentic circulation. This encourages the agent to start out with a plan and replace it after the required variety of steps. From my expertise, this reflection may be very useful for sustaining concentrate on the issue and adjusting actions to remain aligned with the preliminary plan and aim. I undoubtedly advocate utilizing it in circumstances when complicated reasoning is required.

Setting it up is as simple as specifying planning_interval = 3 for the code agent.

agent = CodeAgent(
    mannequin=mannequin,
    instruments=[calculate_simple_growth_metrics],
    max_steps=20,
    additional_authorized_imports=["pandas", "numpy", "matplotlib.*", "plotly.*"],
    verbosity_level = 2, 
    planning_interval = 3,
    managed_agents = [change_log_agent]
)

That’s it. Then, the agent offers reflections beginning with eager about the preliminary plan.

────────────────────────── Preliminary plan ──────────────────────────
Listed below are the info I do know and the plan of motion that I'll 
comply with to unravel the duty:
```
## 1. Information survey

### 1.1. Information given within the job
- We've got a pandas dataframe `df` exhibiting income by section, for 
two time factors: earlier than (April 2025) and after (Could 2025).
- The dataframe columns embrace:
  - Dimensions: `nation`, `maturity`, `country_maturity`, 
`country_maturity_combined`
  - Metrics: `earlier than` (income in April 2025), `after` (income in
Could 2025)
- The duty is to know the modifications in income (after vs 
earlier than) throughout totally different segments.
- Key directions and instruments offered:
  - Determine all dimensions besides earlier than/after for segmentation.
  - Analyze every dimension independently utilizing 
`calculate_simple_growth_metrics`.
  - Filter segments with outsized influence on KPI change (absolute 
normed influence > 1.25).
  - Study mixtures of dimensions if a number of dimensions have
vital segments.
  - Summarize vital modifications and have interaction `change_log_agent` 
for contextual causes.
  - Present a ultimate government abstract together with top-line modifications 
and segment-level detailed impacts.
- Dataset snippet reveals segments combining nations (`France`, 
`UK`, `Germany`, `Italy`, `Spain`, `different`) and maturity standing 
(`new`, `current`).
- The mixed segments are uniquely recognized in columns 
`country_maturity` and `country_maturity_combined`.

### 1.2. Information to lookup
- Definitions or descriptions of the segments if unclear (e.g., 
what defines `new` vs `current` maturity).
  - Probably not necessary to proceed, however could possibly be requested from 
enterprise documentation or change log.
- Extra particulars on the change log (accessible through 
`change_log_agent`) that might present possible causes for income
modifications.
- Affirmation on dealing with mixed dimension splits - how precisely
`country_maturity_combined` is shaped and must be interpreted in
mixed dimension evaluation.
- Knowledge dictionary or description of metrics if any further KPI 
apart from income is related (unlikely given information).
- Dates affirm interval of research: April 2025 (earlier than) and Could 
2025 (after). No must look these up since given.

### 1.3. Information to derive
- Determine all dimension columns accessible for segmentation:
  - By excluding 'earlier than' and 'after', seemingly candidates are 
`nation`, `maturity`, `country_maturity`, and 
`country_maturity_combined`.
- For every dimension, calculate change metrics utilizing the given 
instrument:
  - Absolute and relative distinction in income per section.
  - Affect, section share earlier than, and normed influence for every 
section.
- Determine which segments have outsized influence on KPI change 
(|impact_norm| > 1.25).
- If a number of dimensions have vital segments, mix 
dimensions (e.g., nation + maturity) and reanalyze.
- Decide if mixed dimension splits present significant 
differentiation or not, primarily based on delta price and impact_norm 
consistency.
- Summarize path and magnitude of KPI modifications at top-line 
stage (combination income earlier than and after).
- Determine high segments driving optimistic and unfavourable modifications 
primarily based on ordered absolute absolute_change.
- Collect contextual insights from the change log agent relating to 
possible causes tied to vital segments and the Could 2025 vs 
April 2025 interval.

## 2. Plan

1. Determine all dimension columns current within the dataframe by 
itemizing columns and excluding 'earlier than' and 'after'.
2. For every dimension recognized (`nation`, `maturity`, 
`country_maturity`, `country_maturity_combined`):
   - Use `calculate_simple_growth_metrics` on the complete dataframe 
grouped by that dimension.
   - Extract segments with calculated metrics together with 
impact_norm.
3. Combination outcomes from all single-dimension analyses and filter
segments the place |impact_norm| > 1.25.
4. Decide which dimensions these vital segments belong 
to.
5. If a couple of dimension is represented in these vital 
segments, analyze the mixed dimension shaped by these 
dimensions (for instance, mixture of `nation` and `maturity` 
or use current mixed dimension columns).
6. Repeat metric calculation utilizing 
`calculate_simple_growth_metrics` on the mixed dimension.
7. Study if the mixed dimension splits create significant 
differentiation - if all subsegments present shut difference_rate 
and impact_norm, exclude the break up.
8. Put together a abstract of serious modifications:
   - Prime-line KPIs earlier than and after (absolute and relative 
modifications).
   - Listing of impactful segments sorted by absolute absolute_change
that influenced total income.
9. Present the record of segments with particulars (values earlier than, 
after, absolute and relative change, share earlier than, influence, 
impact_norm).
10. Utilizing this summarized info, question `change_log_agent` 
with full context:
    - Embody vital segments, their relative modifications, and 
durations (April to Could 2025).
11. Course of the agent's response to establish 1-3 foremost possible 
causes of the KPI modifications.
12. Draft government abstract commentary:
    - Excessive-level overview of what occurred and why, primarily based on log 
data.
    - Detailed abstract together with top-line modifications and 
segment-level metrics influence.
13. Ship the ultimate reply utilizing `final_answer` instrument containing 
the above government abstract and data-driven insights.

Then, after every three steps, the agent revisits and updates the plan. 

────────────────────────── Up to date plan ──────────────────────────
I nonetheless want to unravel the duty I used to be given:
```

Here's a pandas dataframe exhibiting the income by section, 
evaluating values earlier than (April 2025) and after (Could 2025). 

You are a senior and skilled information analyst. Your job shall be 
perceive the modifications to the income (after vs earlier than) in 
totally different segments 
and supply government abstract.

<... repeating the complete preliminary job ...>
```

Listed below are the info I do know and my new/up to date plan of motion to 
remedy the duty:
```
## 1. Up to date info survey

### 1.1. Information given within the job
- We've got a pandas dataframe with income by section, exhibiting 
values "earlier than" (April 2025) and "after" (Could 2025).
- Columns within the dataframe embrace a number of dimensions and the 
"earlier than" and "after" income values.
- The aim is to know income modifications by section and supply
an government abstract.
- Steerage and guidelines about  analyze and interpret outcomes 
from the `calculate_simple_growth_metrics` instrument are offered.
- The dataframe incorporates columns: nation, maturity, 
country_maturity, country_maturity_combined, earlier than, after.

### 1.2. Information that we now have discovered
- The size to investigate are: nation, maturity, 
country_maturity, and country_maturity_combined.
- Analyzed income modifications by dimension.
- Solely the "new" maturity section has vital influence 
(impact_norm=1.96 > 1.25), with a big unfavourable income change (~
-70.6%).
- Within the mixed section "country_maturity," the "new" segments 
throughout nations (Spain_new, UK_new, Germany_new, France_new, 
Italy_new, other_new) all have outsized unfavourable impacts with 
impact_norm values all above 1.9.
- The mature/current segments in these nations have smaller 
normed impacts under 1.25.
- Nation-level and maturity-level section dimension alone are 
much less revealing than the mixed nation+maturity section 
dimension which highlights the brand new segments as strongly impactful.
- Complete income dropped considerably from earlier than to after, principally
pushed by new segments shrinking drastically.

### 1.3. Information nonetheless to lookup
- Whether or not splitting the information by further dimensions past 
nation and maturity (e.g., country_maturity_combined) explains 
additional heterogeneous impacts or if the sample is uniform.
- Clarification/context from change log about what precipitated the key 
drop predominantly in new segments in all nations.
- Confirming whether or not any nation throughout the new section behaved 
in a different way or mitigated losses.

### 1.4. Information nonetheless to derive
- A concise government abstract describing the top-level income 
change and figuring out which segments clarify the declines.
- Clarification involving the change log agent with abstract of 
possible causes for these outsized reductions in income within the 
new segments throughout nations for April-Could 2025.

## 2. Plan

### 2.1. Confirm if including the extra dimension 
'country_maturity_combined' splits the impactful "new" segments 
into subsegments with considerably totally different impacts or if the 
change charges and normed impacts are comparatively homogeneous. If 
homogeneous, we don't achieve deeper perception and will disregard 
additional splitting.

### 2.2. Summarize all vital segments recognized with 
outsized impact_norm ≥ 1.25, together with their earlier than and after 
values, absolute and relative modifications, section shares earlier than, 
influence, and normalized influence, ordered by absolute worth of the 
change.

### 2.3. Question the change_log_agent with the complete context: 
vital segments are the brand new country_maturity segments with 
giant unfavourable modifications (~ -70%), timeframe April 2025 to Could 2025,
and request high 1-3 most possible causes for the KPI income drop 
in these segments.

### 2.4. Based mostly on the change log agent's response, synthesize a 
3-5 sentence high-level commentary explaining what occurred 
broadly and why.

### 2.5. Draft an in depth government abstract together with:
- Complete income earlier than and after in human-readable format with 
absolute and relative change.
- A listing of serious segments driving these modifications, so as 
by absolute influence, with detailed numbers (earlier than, after, absolute
and relative change, section share earlier than, influence, normed influence).

### 2.6. Use the `final_answer` instrument to supply the finalized 
government abstract report.

I actually like how the agent is inspired to reiterate on the preliminary job and keep targeted on the primary downside. Common reflection like that is useful in actual life as nicely, as groups typically get slowed down within the course of and lose sight of the why behind what they’re doing. It’s fairly cool to see managerial greatest practices being built-in into agentic frameworks.

That’s it! We’ve constructed a code agent able to analysing KPI modifications for easy metrics and explored all the important thing nuances of the method.

You could find the entire code and execution logs on GitHub.

Abstract

We’ve experimented rather a lot with code brokers and are actually prepared to attract conclusions. For our experiments, we used the HuggingFace smolagents framework for code brokers — a really useful toolset that provides: 

• simple integration with totally different LLMs (from native fashions through Ollama to public suppliers like Anthropic or OpenAI),
• excellent logging that makes it simple to know the entire thought means of the agent and debug points,
• means to construct complicated programs leveraging multi-AI agent setups or planning options with out a lot effort.

Whereas smolagents is presently my favorite agentic framework, it has its limitations: 

• It could lack flexibility at instances. For instance, I needed to modify the immediate straight within the supply code to get the behaviour I needed.
• It solely helps hierarchical multi-agent set-up (the place one supervisor can delegate duties to different brokers), however doesn’t cowl sequential workflow or consensual decision-making processes.
• There’s no help for long-term reminiscence out of the field, that means you’re ranging from scratch with each job.

Thank you a large number for studying this text. I hope this text was insightful for you.

Reference

This text is impressed by the “Constructing Code Brokers with Hugging Face smolagents” brief course by DeepLearning.AI.