Pass or extract files explicitly to AgentCore Code Interpreter using Strands Agents' built-in tools

from bedrock_agentcore.tools.code_interpreter_client import CodeInterpreter

code_client = CodeInterpreter('us-east-1')
code_client.start()

# Explicitly write files
files_to_create = [
    {"path": "data.csv", "text": data_content},
    {"path": "stats.py", "text": stats_content}
]
code_client.invoke("writeFiles", {"content": files_to_create})

# Execute code
code_client.invoke("executeCode", {"code": "exec(open('stats.py').read())", "language": "python"})

from strands_tools.code_interpreter import AgentCoreCodeInterpreter
from strands_tools.code_interpreter.models import WriteFilesAction, FileContent

code_interpreter = AgentCoreCodeInterpreter(region="us-west-2")

# Explicitly write files (not relying on the agent)
code_interpreter.write_files(WriteFilesAction(
    type="writeFiles",  # Required field
    content=[
        FileContent(path="data.csv", text=data_content),
        FileContent(path="stats.py", text=stats_content)
    ]
))

uv init
uv add strands-agents strands-agents-tools bedrock-agentcore bedrock-agentcore-starter-toolkit aws-opentelemetry-distro

name,age,score,department
田中太郎,28,85,営業
佐藤花子,34,92,開発
鈴木一郎,25,78,マーケティング
高橋美咲,31,88,開発
伊藤健太,29,95,営業
渡辺真理,27,72,マーケティング
山本大輔,33,91,開発
中村愛,26,83,営業
小林拓也,30,87,開発
加藤由美,32,79,マーケティング

import pandas as pd

# Load data
df = pd.read_csv('data.csv')

print("=== Data Analysis Results ===\n")

# Basic statistics
print("【Basic Statistics】")
print(f"Number of records: {len(df)}")
print(f"Average age: {df['age'].mean():.1f} years")
print(f"Average score: {df['score'].mean():.1f} points")

# Department statistics
print("\n【Average Score by Department】")
dept_stats = df.groupby('department')['score'].agg(['mean', 'count'])
for dept, row in dept_stats.iterrows():
    print(f"  {dept}: {row['mean']:.1f} points ({int(row['count'])} people)")

# Highest score
top = df.loc[df['score'].idxmax()]
print(f"\n【Highest Score】")
print(f"  {top['name']} ({top['department']}): {top['score']} points")

from strands import Agent
from strands.models import BedrockModel
from strands_tools.code_interpreter import AgentCoreCodeInterpreter
from strands_tools.code_interpreter.models import WriteFilesAction, FileContent

# Model settings
model = BedrockModel(
    model_id="us.anthropic.claude-sonnet-4-5-20250929-v1:0",
    region_name="us-west-2"
)

code_interpreter = AgentCoreCodeInterpreter(region="us-west-2")

# Read local files
with open("data.csv", "r") as f:
    data_content = f.read()
with open("stats.py", "r") as f:
    stats_content = f.read()

print("=== Checking File Contents ===")
print(f"data.csv: {len(data_content)} bytes")
print(f"stats.py: {len(stats_content)} bytes")

# Explicitly write files (not relying on the agent)
print("\n=== Writing Files to Sandbox... ===")
result = code_interpreter.write_files(WriteFilesAction(
    type="writeFiles",  # Required field
    content=[
        FileContent(path="data.csv", text=data_content),
        FileContent(path="stats.py", text=stats_content)
    ]
))
print(f"Write result: {result}")

# Agent only uses executeCode
print("\n=== Running Agent ===")
agent = Agent(
    model=model,
    tools=[code_interpreter.code_interpreter],
    system_prompt="Code Interpreter already has data.csv and stats.py. Please execute the code using executeCode."
)

response = agent("Run stats.py and tell me the results")
print(f"\n=== Results ===")
print(response)

uv run test_file_upload.py

=== Checking File Contents ===
data.csv: 180 bytes
stats.py: 559 bytes

=== Writing Files to Sandbox... ===
Write result: {'status': 'success', 'content': [{'text': "[{'type': 'text', 'text': 'Successfully wrote all 2 files'}]"}]}

=== Running Agent ===
I'll run stats.py.
Tool #1: code_interpreter
I've run stats.py. Here are the results:

## Data Analysis Results

**【Basic Statistics】**
- Number of records: 10
- Average age: 29.5 years
- Average score: 85.0 points

**【Average Score by Department】**
- Marketing: 76.3 points (3 people)
- Sales: 87.7 points (3 people)
- Development: 89.5 points (4 people)

**【Highest Score】**
- Ito Kenta (Sales): 95 points

The Development department has the highest average score (89.5 points), and Ito Kenta from the Sales department achieved the highest score of 95 points.

from strands import Agent
from strands.models import BedrockModel
from strands_tools.code_interpreter import AgentCoreCodeInterpreter
from strands_tools.code_interpreter.models import (
    WriteFilesAction,
    FileContent,
    ReadFilesAction,
    ListFilesAction,
    ExecuteCodeAction,
)
import base64

# Model settings
model = BedrockModel(
    model_id="us.anthropic.claude-sonnet-4-5-20250929-v1:0",
    region_name="us-west-2"
)

code_interpreter = AgentCoreCodeInterpreter(region="us-west-2")

# Read local files
with open("data.csv", "r") as f:
    data_content = f.read()

print("=== Step 1: Writing CSV File to Sandbox ===")
result = code_interpreter.write_files(WriteFilesAction(
    type="writeFiles",
    content=[
        FileContent(path="data.csv", text=data_content),
    ]
))
print(f"Write result: {result}")

print("\n=== Step 2: Executing Code to Generate Graphs ===")
chart_code = """
import pandas as pd
import matplotlib.pyplot as plt
import matplotlib
matplotlib.use('Agg')  # Draw without GUI

# Load data
df = pd.read_csv('data.csv')

# Map department names to English
dept_mapping = {
    'マーケティング': 'Marketing',
    '営業': 'Sales',
    '開発': 'Development'
}
df['department_en'] = df['department'].map(dept_mapping)

# Calculate average scores by department
dept_scores = df.groupby('department_en')['score'].mean().sort_values(ascending=True)

# Create graph
plt.figure(figsize=(10, 6))
bars = plt.barh(dept_scores.index, dept_scores.values, color=['#3498db', '#2ecc71', '#e74c3c'])

# Display values next to bars
for bar, value in zip(bars, dept_scores.values):
    plt.text(value + 0.5, bar.get_y() + bar.get_height()/2, f'{value:.1f}',
             va='center', fontsize=12)

plt.xlabel('Average Score', fontsize=12)
plt.ylabel('Department', fontsize=12)
plt.title('Average Score by Department', fontsize=14, fontweight='bold')
plt.xlim(0, 100)
plt.tight_layout()

# Save as image
plt.savefig('chart.png', dpi=100, bbox_inches='tight')
plt.close()

print("chart.png generated")

# Also generate CSV report
report = df.groupby('department').agg({
    'score': ['mean', 'min', 'max', 'count']
}).round(1)
report.columns = report.columns.droplevel(0)  # Remove MultiIndex
report = report.reset_index()  # Make department a column
report.to_csv('report.csv', index=False)
print("report.csv generated")
"""

result = code_interpreter.execute_code(ExecuteCodeAction(
    type="executeCode",
    code=chart_code,
    language="python"
))
print(f"Execution result: {result}")

print("\n=== Step 3: Checking File List in Sandbox ===")
result = code_interpreter.list_files(ListFilesAction(
    type="listFiles",
    path="."
))
print(f"File list: {result}")

print("\n=== Step 4: Reading Generated Files ===")
result = code_interpreter.read_files(ReadFilesAction(
    type="readFiles",
    paths=["report.csv", "chart.png"]
))
print(f"Read result: {result}")

# Parse results and save locally
print("\n=== Step 5: Saving Files Locally ===")

# Parse read_files results
import ast
content_str = result['content'][0]['text']
resources = ast.literal_eval(content_str)

for resource in resources:
    if resource['type'] == 'resource':
        res = resource['resource']
        uri = res['uri']
        mime_type = res['mimeType']
        filename = uri.split('/')[-1]

        print(f"\nFile: {filename} (MIME: {mime_type})")

        if 'text' in res:
            # For text files
            with open(f"output_{filename}", "w") as f:
                f.write(res['text'])
            print(f"  → Saved to output_{filename}")
            print(f"  Content preview:\n{res['text'][:200]}")

        elif 'blob' in res:
            # For binary files
            blob_data = res['blob']
            with open(f"output_{filename}", "wb") as f:
                f.write(blob_data)
            print(f"  → Saved to output_{filename} ({len(blob_data)} bytes)")

print("\n=== Complete ===")

uv run test_read_files.py

{'status': 'success', 'content': [{'text': "[
  {'type': 'resource_link', 'mimeType': 'image/png', 'uri': 'file:///chart.png', 'name': 'chart.png', 'description': 'File'},
  {'type': 'resource_link', 'mimeType': 'text/csv', 'uri': 'file:///data.csv', 'name': 'data.csv', 'description': 'File'},
  {'type': 'resource_link', 'mimeType': 'text/csv', 'uri': 'file:///report.csv', 'name': 'report.csv', 'description': 'File'}
]"}]}

import ast

content_str = result['content'][0]['text']
resources = ast.literal_eval(content_str)

for resource in resources:
    if resource['type'] == 'resource':
        res = resource['resource']
        uri = res['uri']
        mime_type = res['mimeType']
        filename = uri.split('/')[-1]

        if 'text' in res:
            # For text files
            with open(f"output_{filename}", "w") as f:
                f.write(res['text'])

        elif 'blob' in res:
            # For binary files
            blob_data = res['blob']
            with open(f"output_{filename}", "wb") as f:
                f.write(blob_data)

department,mean,min,max,count
マーケティング,76.3,72,79,3
営業,87.7,83,95,3
開発,89.5,87,92,4

from bedrock_agentcore.runtime import BedrockAgentCoreApp
from strands import Agent
from strands.models import BedrockModel
from strands_tools.code_interpreter import AgentCoreCodeInterpreter
from strands_tools.code_interpreter.models import WriteFilesAction, FileContent

app = BedrockAgentCoreApp()

# Sample data for analysis
SAMPLE_DATA = """name,age,score,department
田中太郎,28,85,営業
佐藤花子,34,92,開発
鈴木一郎,25,78,マーケティング
高橋美咲,31,88,開発
伊藤健太,29,95,営業
渡辺真理,27,72,マーケティング
山本大輔,33,91,開発
中村愛,26,83,営業
小林拓也,30,87,開発
加藤由美,32,79,マーケティング"""

@app.entrypoint
def main(payload):
    model = BedrockModel(
        model_id="us.anthropic.claude-sonnet-4-5-20250929-v1:0",
        region_name="us-west-2"
    )

    code_interpreter = AgentCoreCodeInterpreter(region="us-west-2")

    # Write CSV file to sandbox
    code_interpreter.write_files(WriteFilesAction(
        type="writeFiles",
        content=[FileContent(path="data.csv", text=SAMPLE_DATA)]
    ))

    # Give the agent ambiguous instructions
    agent = Agent(
        model=model,
        tools=[code_interpreter.code_interpreter],
        system_prompt="You are a data analyst. data.csv is in the Code Interpreter. Write your own code to analyze the user's questions."
    )

    # Ambiguous instruction: Let the agent decide how to analyze
    response = agent("Analyze data.csv and tell me three interesting insights")
    return response

if __name__ == "__main__":
    app.run()

uv run agentcore configure --entrypoint test_agent_analysis.py
# Choose mostly automatic creation options, select Direct Code Deploy

uv run agentcore launch 

uv run agentcore invoke '{"prompt": ""}'

Response:
After analyzing data.csv, I found the following **three interesting insights**:

## 📊 Insight 1: The Development Department is Overwhelmingly Excellent and Stable
- **Development Department Average Score**: 89.5 points (highest among all departments)
- **Score Stability**: Standard deviation of just 2.38, the most consistent performance
- All 4 people in the Development Department scored 87 points or higher, maintaining a high level across the team

In contrast, the Marketing Department has an average of 76.3 points, about 13 points lower, showing the most room for improvement.

## 📈 Insight 2: There is a moderate positive correlation (0.562) between age and score
- **Young employees (~27 years)**: Average 77.7 points
- **Mid-career (28-30 years)**: Average 89.0 points (highest!)
- **Veteran (31+ years)**: Average 87.5 points

Surprisingly, **the 28-30 year mid-career group shows the highest performance**. This suggests a balance between experience and energy, indicating that the relationship is not simply "higher age equals better performance."

## 🏆 Insight 3: Top performers are older and concentrated in the Development Department
- **Characteristics of the top 30% (3 people)**:
  - Average age: 32.0 years (higher than the overall average of 29.5 years)
  - Average score: 92.7 points
  - **Department distribution**: 2 in Development, 1 in Sales (0 in Marketing)

Two-thirds of top performers belong to the Development Department, suggesting that the Development Department either attracts talented people or has an excellent development environment. It's notable that there are no top performers from the Marketing Department.

Event 3: gen_ai.choice (body)
"code": "import pandas as pd\nimport numpy as np\nimport matplotlib.pyplot as plt\nimport seaborn as sns\n\n# データの読み込み\ndf = pd.read_csv('data.csv')\n..."

Event 7: gen_ai.tool.message (body)
"ModuleNotFoundError: No module named 'seaborn'"

Event 9: gen_ai.choice (body)
"code": "import pandas as pd\nimport numpy as np\nimport matplotlib.pyplot as plt\n\n# データの読み込み\ndf = pd.read_csv('data.csv')\n\n# 基本情報の確認\nprint('データの形状:', df.shape)\nprint('\n最初の数行:')\nprint(df.head(10))\n..."

Event 16: gen_ai.tool.message (body)
データの形状: (10, 4)

最初の数行:
   name  age  score department
0  田中太郎   28     85         営業
1  佐藤花子   34     92         開発
...

基本統計量:
            age      score
mean   29.50000  85.000000
std     3.02765   7.118052

Event 18: gen_ai.choice (body)
"code": "# 洞察1: 部署別の分析\ndept_score = df.groupby('department')['score'].agg(['mean', 'count'])\n\n# 洞察2: 年齢とスコアの関係\ncorrelation = df['age'].corr(df['score'])\ndf['age_group'] = pd.cut(df['age'], bins=[0, 27, 30, 35], labels=['若手(~27歳)', '中堅(28-30歳)', 'ベテラン(31歳~)'])\n\n# 洞察3: トップパフォーマー分析\ntop_threshold = df['score'].quantile(0.7)\n..."

Event 28: gen_ai.tool.message (body)
=== 洞察1: 部署別パフォーマンス分析 ===
部署別の平均スコア:
                 mean  count
department                  
開発          89.500000      4
営業          87.666667      3
マーケティング     76.333333      3

=== 洞察2: 年齢とスコアの相関分析 ===
年齢とスコアの相関係数: 0.562
年齢層別の平均スコア:
若手(~27歳)    77.666667
中堅(28-30歳)  89.000000
ベテラン(31歳~)  87.500000

=== 洞察3: トップパフォーマー分析 ===
上位30%のスコア閾値: 88.9
トップパフォーマーの特徴:
人数: 3人
平均年齢: 32.0歳
平均スコア: 92.7点
部署別分布: 開発2人、営業1人

Event 30: gen_ai.choice (body)
"code": "# さらに詳細な分析\nfor dept in df['department'].unique():\n    dept_data = df[df['department'] == dept]\n    print(f'  スコアの標準偏差: {dept_data[\"score\"].std():.2f}')\n..."

Event 43: gen_ai.tool.message (body)
=== 追加分析 ===
営業: スコアの標準偏差: 6.43
開発: スコアの標準偏差: 2.38
マーケティング: スコアの標準偏差: 3.79

最高スコア: 95点 - 伊藤健太 (営業, 29歳)
最低スコア: 72点 - 渡辺真理 (マーケティング, 27歳)

Event 48: Invocation completed successfully (41.517s)

data.csvを分析した結果、以下の**3つの興味深い洞察**が見つかりました:

## 📊 洞察1: 開発部署が圧倒的に優秀かつ安定している
- **開発部署の平均スコア**: 89.5点（全部署中トップ）
- **スコアの安定性**: 標準偏差わずか2.38で、最も一貫したパフォーマンス
- 開発部署は4人中全員が87点以上を取得しており、チーム全体として高い水準を維持

一方、マーケティング部署は平均76.3点と約13点も低く、最も改善の余地があります。

## 📈 洞察2: 年齢とスコアに中程度の正の相関（0.562）がある
- **若手（~27歳）**: 平均77.7点
- **中堅（28-30歳）**: 平均89.0点（最も高い！）
- **ベテラン（31歳~）**: 平均87.5点

意外にも、**28-30歳の中堅層が最も高いパフォーマンス**を示しています。これは経験と活力のバランスが取れた年齢層であることを示唆して
おり、単純な「年齢が高いほど良い」という関係ではないことが分かります。

## 🏆 洞察3: トップパフォーマーは年齢が高く、開発部署に集中
- **上位30%（3人）の特徴**:
  - 平均年齢: 32.0歳（全体平均29.5歳より高い）
  - 平均スコア: 92.7点
  - **部署分布**: 開発2人、営業1人（マーケティング0人）

トップパフォーマーの2/3が開発部署に所属しており、開発部署が優秀人材を集めているか、育成環境が優れている可能性があります。マーケティング部署からはトップパフォーマーが1人も出ていない点は要注意です。