Amazon Bedrock AgentCore engineer Kevin Orellana describes exposing a cloud sandbox through MCP and asking coding agents to generate and execute roughly a thousand workflows across ten domains. Results were stored in SQLite so multiple local agents could record structured findings. The broad test count was a way to discover failure patterns, not a proof of exhaustive coverage.
Examples include a corrupted image round trip caused by MCP encoding, null-byte handling and files that could not be retrieved from the expected directory. Orellana separates interface failures from sandbox failures and compares code execution with browser workloads, where many tabs can exhaust memory. File permissions, persistence, process controls, installed packages and network egress all affect whether an agent can complete its task.
The talk emphasizes testing the actual remote environment rather than assuming local success transfers unchanged. Fuzzy, multi-domain workflows can expose sharp edges that ordinary API expectations miss. The resulting failures need attribution and investigation; an unsuccessful task is not necessarily evidence that the model itself lacks the required capability.
Watch on YouTube
