Code Interpreter

Do NOT use this reference for:

• Interacting with web pages — see browser.md
• Running arbitrary long-lived services — Code Interpreter is for short-lived code execution, not hosting servers
• Shell commands inside your live agent session’s own microVM — that’s InvokeAgentRuntimeCommand, covered in integrate.md

Mental model

Code Interpreter is a managed sandbox, one per session, running in an isolated microVM. Your code can:

• Execute Python, JavaScript, or TypeScript
• Read/write files on a local filesystem (up to 100 MB inline upload, up to 5 GB via S3)
• Make network calls (if internet access is enabled on the resource)
• Use pre-installed libraries (pandas, numpy, scikit-learn, torch, etc. — see docs for the current list)

Sessions are stateful within a session (variables and files persist across execute_code calls in the same session) and ephemeral across sessions (start a new session and the filesystem is clean).

Prerequisites

• Python 3.10+ in your agent environment
• bedrock-agentcore SDK
• IAM permissions for bedrock-agentcore:*CodeInterpreter* actions, scoped to the resource ARN
• Model access if calling via an agent framework (the framework calls a model to decide when to execute code)

IAM policy skeleton:

Check current action names against the docs — the list evolves.

Path A — Strands agent with Code Interpreter (recommended for most)

Install: pip install bedrock-agentcore strands-agents strands-agents-tools

The agent decides when to execute code, starts sessions on demand, and stops them. Under the hood, the tool uses the AWS-managed aws.codeinterpreter.v1 resource — no resource creation needed.

Dropping into an AgentCore Runtime entrypoint:

Path B — Direct SDK for programmatic execution

Use when your code — not an agent — decides what to run. Good for ETL, data transformation, and agent-internal validation.

The context manager handles start/stop. Do not leak sessions.

Language selection — default is Python. For JavaScript/TypeScript, pass language="javascript" or language="typescript" to execute_code (or the runtime setting at session start). See the runtime selection doc for the current supported runtimes.

Path C — Custom Code Interpreter with S3 access

The managed aws.codeinterpreter.v1 resource has no S3 write permissions. For agents that produce artifacts (plots, reports, processed datasets) you want to persist, create a custom Code Interpreter with an execution role that has S3 access.

This is a CreateCodeInterpreter call (SDK/API, not exposed via agentcore CLI at time of writing). The execution role’s trust policy grants bedrock-agentcore.amazonaws.com the ability to assume it, and its permissions policy grants s3:PutObject and related actions on your artifact bucket. Check the docs for the current CreateCodeInterpreter shape and the exact trust policy format.

Same-account S3 rule. The S3 bucket must be in the same AWS account as the Code Interpreter resource. Cross-account buckets are not supported as targets even with the right bucket policy — CreateCodeInterpreter fails with a validation error. If you need the artifacts in another account, replicate from the same-account bucket afterward.

Observability

• CloudWatch logs — stdout/stderr from executed code, plus session lifecycle events
• CloudTrail — every StartCodeInterpreterSession, InvokeCodeInterpreter, StopCodeInterpreterSession call
• Metrics — in AWS/BedrockAgentCore namespace

Pre-installed libraries

The managed Python runtime includes: pandas, numpy, scipy, matplotlib, plotly, scikit-learn, torch, torchvision, statsmodels, and dozens more for data analysis / ML. Check the current list in the docs before telling a user “library X is preinstalled” — the list changes with platform updates.

For libraries not preinstalled, call install_packages(["your-lib==1.2"]) in your session (or !pip install ... via execute_command). Installed packages last only for the session.

Session lifecycle — always close

Default session timeout is 900 seconds (15 min), max 28800 seconds (8 hours). Leaked sessions cost money.

VPC mode

Code Interpreter supports VPC — same pattern as Runtime and Browser (service-linked role, your subnets, your security group). See vpc.md.

Public internet from the sandbox requires a NAT gateway on a private subnet, same as Runtime. Public subnets don’t give Code Interpreter ENIs internet access. If the code needs pip install to reach PyPI, plan for NAT.

Common failures

“Access denied” on StartCodeInterpreterSession: IAM missing the action on the resource ARN. Use aws sts get-caller-identity to confirm which identity you attached the policy to.

“ValidationException: Role does not have access to required S3 buckets”: S3 bucket is in a different account. Move the bucket or replicate from an in-account staging bucket.

Code times out: Default execute timeout is short. Split long jobs into chunks, or use a custom Code Interpreter with extended timeouts. Don’t try to run 30-minute training jobs in Code Interpreter — that’s a SageMaker / Batch job.

“Module not found” despite being listed as preinstalled: The preinstalled list may differ between python and nodejs runtimes. Verify runtime selection and list matches.

Output

• Which path fits (Strands tool vs direct SDK vs custom with S3)
• Working code with session lifecycle handled
• IAM policy scoped to the code-interpreter resource

Quality criteria

• Sessions are always wrapped in a context manager or try/finally — never leaked
• IAM is scoped to code-interpreter/* in the account, not Resource: "*"
• S3 destination buckets are in the same account as the Code Interpreter resource
• Language / runtime selection is explicit when the code isn’t Python