AWS Engineer Kevin Orellana Discusses Stress-Testing Sandboxes with Coding Agents

On October 12, 2026, the AI Engineer channel published the video "1,000 Agent Tasks in a Sandbox: What Breaks When LLMs Run Code," featuring Kevin Orellana, Software Engineer at Amazon Bedrock AgentCore. The talk was recorded at the AI Engineer World's Fair 2026 in San Francisco.

According to the publisher's description and chapters, Orellana presented a workflow where he wrapped AgentCore's code sandbox and cloud browser in an MCP server and had a coding agent write and run 1,000 tests across 10 domains, logging results to SQLite. Listed topics include using coding agents plus MCP to stress-test platforms, why results belong in a database rather than markdown files, distinguishing interface-layer bugs from sandbox-layer limits (such as memory, file permissions, seccomp, and eBPF), and how sandbox limits determine whether an agent can finish its task.

For engineering professionals focused on AI agent infrastructure reliability and security boundaries, this resource provides specific methodological context. Readers may look for discussions on differentiating application-layer logic errors from underlying system resource constraints, as well as the limitations of LLM-generated tests.