---
title: "I got frustrated with AWS, so we built our own sandbox"
description: "At a healthcare startup, an agent workflow built with LangGraph was constrained by context length and by the limited tools exposed to it for data analysis. Changing one node gave the agent access to Amazon Bedrock AgentCore Code Interpreter, allowing it to use pandas and other available data-science libraries in a sandbox. It could inspect files and outputs, write and run code, and adjust its analysis iteratively while keeping working data and intermediate artifacts in the workspace rather than sending every row through the conversation. Although the experiment demonstrated the usefulness of this controlled autonomy, regional availability forced the author to use a US region from Malaysia, motivating BoxCompute, an infrastructure project intended to provide agent workspaces with control over execution, resources, output collection, and cleanup."
---

# I got frustrated with AWS, so we built our own sandbox

[Boxcompute](https://yomu.fyi/company/boxcompute) · Farhan Helmy · Sep 16, 2026

**Type:** Explainer

## Summary

At a healthcare startup, an agent workflow built with LangGraph was constrained by context length and by the limited tools exposed to it for data analysis. Changing one node gave the agent access to Amazon Bedrock AgentCore Code Interpreter, allowing it to use pandas and other available data-science libraries in a sandbox. It could inspect files and outputs, write and run code, and adjust its analysis iteratively while keeping working data and intermediate artifacts in the workspace rather than sending every row through the conversation. Although the experiment demonstrated the usefulness of this controlled autonomy, regional availability forced the author to use a US region from Malaysia, motivating BoxCompute, an infrastructure project intended to provide agent workspaces with control over execution, resources, output collection, and cleanup.

## Context

The work grew from a LangGraph workflow whose roles included routing questions, planning analyses, querying data, and interpreting results. The author was investigating how a workspace could let agents perform intermediate data-analysis steps without exposing every transformation as a separate tool or placing all working data in the conversation.

## Approach / What changed

A single workflow node was changed to give the agent Amazon Bedrock AgentCore Code Interpreter access. In the sandbox, the agent could inspect files, write and execute code, examine outputs, retry with adjusted analysis, and retain working data and intermediate artifacts. BoxCompute extends that workspace idea by owning environment startup, resource management, output collection, and cleanup.

## Takeaways

- A sandbox lets an agent choose intermediate data operations, write and run code, inspect results, and adjust its next step without requiring every transformation to be predefined as a tool.
- Keeping datasets in the workspace and returning focused outputs reduces the need to place every row in the model’s conversation, although the agent still needs enough context to decide what to do next.
- Regional availability of Amazon Bedrock AgentCore Code Interpreter led the author, working from Malaysia, to build BoxCompute and control the execution environment directly.

## FAQ

### How did the sandbox change the agent’s data-analysis workflow?

The agent could inspect files, write and execute analysis code, review outputs, and retry with adjusted code. It could use pandas and other available data-science libraries without requiring each possible transformation to be exposed as a separate tool.

### How did the workspace help with context-window limits?

The dataset and intermediate artifacts remained in the workspace while code produced summaries, filtered rows, or calculated statistics. The agent could inspect smaller outputs instead of sending every row through the conversation, though it still needed enough context to choose its next action.

### Why was BoxCompute built instead of continuing with the managed interpreter?

Amazon Bedrock AgentCore Code Interpreter was unavailable in the region the author wanted at the time, so the author used a US region from Malaysia. BoxCompute was built from scratch to provide more control over where workspaces run and to manage their environments, resources, outputs, and cleanup.

**Tags:** [AI Agents](https://yomu.fyi/topic/ai-agents), [LangGraph](https://yomu.fyi/topic/langgraph), [Sandboxing](https://yomu.fyi/topic/sandboxing)

- Source: [Boxcompute](https://boxcompute.ai/blog/why-we-built-boxcompute)
- Source URL: https://boxcompute.ai/blog/why-we-built-boxcompute
- Ingested by Yomu: 2026-09-18T03:34:59.930Z

[Read original post](https://boxcompute.ai/blog/why-we-built-boxcompute)