Migration: Anthropic Code Execution Tool to AgentCore Code Interpreter

Overview

This snippet shows how to migrate from the Anthropic API Code Execution tool to Amazon Bedrock + AgentCore Code Interpreter. It provides side-by-side working examples: the original Anthropic approach (one API call, everything handled server-side) and multiple equivalent implementations on AWS.

All examples (00-04, 06) use the same prompt (a simple math calculation) — “Calculate the mean and standard deviation of [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]” — so you can directly compare the code and output across approaches.

Key Architectural Difference

Anthropic	AWS (Bedrock + AgentCore)
One API call does everything	Two services composed together
No agentic loop needed	Orchestration layer required (or use Strands)
No session management	Explicit session start/stop lifecycle

Technologies

Amazon Bedrock (Claude via Converse API and InvokeModel API)
Amazon Bedrock AgentCore Code Interpreter
Strands Agents SDK
Claude Code CLI (in sandbox)
Bedrock Prompt Caching
Anthropic API (reference only)
Python 3.10+

Difficulty

Intermediate

Sample Files

File	Approach	Description
`00_anthropic_code_execution.py`	Anthropic API	BEFORE — The original. One API call, Claude handles code execution server-side. No loop, no session management.
`01_converse_api_with_code_interpreter.py`	Converse API + AgentCore	Custom agentic loop using the AWS-recommended Converse API. Model-agnostic format that works across Claude, Nova, etc.
`02_invoke_model_with_code_interpreter.py`	InvokeModel API + AgentCore	Custom agentic loop using native Anthropic Messages format. Easiest to port from existing Anthropic code.
`03_strands_agent_with_code_interpreter.py`	Strands SDK + AgentCore	RECOMMENDED — Strands handles the agentic loop automatically. Closest to the Anthropic “one call” simplicity.
`04_code_interpreter_highlevel_api.py`	AgentCore SDK (direct)	Execute code directly without a model. Useful for batch jobs, testing, or when you already know what code to run.
`05_claude_code_in_code_interpreter/`	Claude Code in sandbox	Run Claude Code CLI inside a Code Interpreter sandbox. Requires a custom Code Interpreter with PUBLIC network mode and an IAM execution role. See 05 README.
`06_invoke_model_with_caching.py`	InvokeModel + Prompt Caching	Demonstrates Bedrock prompt caching to reduce cost and latency in agentic loops. Uses a regional model ID (required for caching).
`07_programmatic_tool_calling.py`	Programmatic Tool Calling	Demonstrates the programmatic tool calling pattern on Bedrock: pre-loads tools into the sandbox so the model can call multiple tools in a single code execution — reducing round-trips and token usage.

Choosing the Right Approach

Migrating existing Anthropic code? Start with 02 (InvokeModel) — the request format is nearly identical.
New project or want simplicity? Use 03 (Strands) — one call does everything, just like Anthropic.
Need model-agnostic code? Use 01 (Converse) — works with Claude, Nova, Cohere, etc.
Don’t need a model? Use 04 (direct) — just run code in a sandbox.
Want a full coding agent in a sandbox? Use 05 (Claude Code) — runs the full Claude Code CLI inside Code Interpreter.
Optimizing cost for agentic loops? Use 06 (caching) — caches system prompt and tools across repeated calls.
Calling many tools per turn? Use 07 (programmatic) — model writes code that calls all tools at once, reducing round-trips.

Network Mode Impact

Most examples (01-04, 06, 07) use the default sandbox (aws.codeinterpreter.v1) — no network access, no custom setup. This is sufficient for computation, data analysis, and demos with static data.

For real tool integrations (APIs, databases, AWS services), use a custom Code Interpreter with PUBLIC network mode and an IAM execution role. This lets your pre-loaded functions make real HTTP/SDK calls directly from inside the sandbox. See 05_claude_code_in_code_interpreter/ for setup.

Network Mode	Use Case	Setup
Default (no network)	Computation, data analysis, demos	None — works out of the box
PUBLIC + IAM role	Real API calls, DB queries, AWS SDK	Custom Code Interpreter (setup guide)

Model ID Notes

Scripts 01-04 use global.anthropic.claude-sonnet-4-6 (cross-region routing for lowest latency).
Script 06 uses us.anthropic.claude-sonnet-4-5-20250929-v1:0 (regional ID, required for prompt caching — cross-region IDs route to different regions, preventing cache hits).

Prerequisites

AWS Account with Amazon Bedrock access
Claude model enabled in the Bedrock console (us-west-2)
IAM credentials with AgentCore and Bedrock permissions
Python 3.10+
AWS CLI configured (aws configure or SSO)

Setup

# Navigate to the snippet directory
cd snippets/migration-from-code-execution-tool-to-agentcore-code-interpreter

# Create and activate virtual environment
python3 -m venv .venv
source .venv/bin/activate

# Install dependencies
pip install -r requirements.txt

Run the examples

# Recommended: Strands SDK (simplest, closest to Anthropic experience)
python 03_strands_agent_with_code_interpreter.py

# Converse API with custom agentic loop
python 01_converse_api_with_code_interpreter.py

# InvokeModel API with custom agentic loop (native Anthropic format)
python 02_invoke_model_with_code_interpreter.py

# Direct code execution (no model)
python 04_code_interpreter_highlevel_api.py

# InvokeModel with prompt caching
python 06_invoke_model_with_caching.py

# Programmatic tool calling (multiple tools in one code execution)
python 07_programmatic_tool_calling.py

# Claude Code in sandbox (see 05_claude_code_in_code_interpreter/README.md)
cd 05_claude_code_in_code_interpreter
python setup_infrastructure.py  # one-time
python run.py <code-interpreter-id> "Your prompt"

# Anthropic reference (requires ANTHROPIC_API_KEY)
# export ANTHROPIC_API_KEY=your_key
# python 00_anthropic_code_execution.py

Security

See CONTRIBUTING for more information.

License

This library is licensed under the MIT-0 License. See the LICENSE file.

View Source

View on GitHub