Markdown source

1,000 Agent Tasks in a Sandbox: What Breaks When LLMs Write and Run Code

Conference Context

Session Description

We ran 1,000 automated tasks through a production code interpreter sandbox — file I/O, package installs, data analysis, ML training, binary downloads, multi-language execution — and tracked every failure. 88% passed. The other 12% revealed 18 distinct failure modes that no unit test would catch: binary encoding corruption in the transport layer, null bytes silently truncating file downloads, pip blocked by network isolation with no useful error, and path traversal inputs accepted without validation. This talk walks through the experiment design, the findings ranked by severity, and what we changed. If you are building or operating sandboxed execution for AI agents, these are the bugs waiting for your customers to find first.

Media Evidence

No related AI Engineer channel video found yet.

Evidence Graph

This evidence graph is generated from currently linked source material: official schedule text, related video pages, cached transcripts, visible slide text, dense/reconstructed slide pages, and AI slide-classification audits.

Media Signals

No linked video, transcript, or slide source has been attached yet.

Agent Reading Notes

Use these signals to refine the synopsis, topic links, people/company context, and method notes. If a source is a related external video rather than an exact official recording, keep it framed as supporting evidence.

Transcript Status

No official session recording transcript was found by exact title match on the AI Engineer YouTube channel during this run.

People

Notes

Synthesis

Synthesized Breakdown

1,000 Agent Tasks in a Sandbox: What Breaks When LLMs Write and Run Code ## Conference Context - Date/time: 2026-06-30 · 2:25pm-2:45pm - Track/room: Sandbox & Platform Engineering · Track 1 - Speaker(s): Kevin Orellana - Session type/status: session · confirmed - Track: Sandbox & Platform Engineering - Room: Track 1 - Session type: session - Status: confirmed ## Session Description We ran 1,000 automated tasks through a production code interpreter sandbox — file I/O, package installs, data analysis, ML training, binary downloads, multi-language execution — and tracked every failure. 88% passed. The other 12% revealed 18 distinct failure modes that no unit test would catch: binary encoding corruption in the transport layer, null bytes silently truncating file downloads, pip blocked by network isolation with no useful error, and path traversal inputs accepted without validation. This talk walks through the experiment design, the findings ranked by severity, and what we changed.

Speaker And Company Context

Topics Covered

Derived Links And Source Material

Novel Concepts / Clever Methods

Evidence Boundary

This synthesis is based on the official schedule and linked source pages. It should be revisited when exact session recordings or transcript-backed secondary sources are available.