DigitalOcean Targets Silent Agent Failures

This closes one of the biggest gaps in agent operations: connecting real-world failures to repeatable pre-release tests. The approach is promising, but preview limitations and delayed analysis mean it complements rather than replaces robust state-based testing.

• –Signals classifies interaction, execution, and infrastructure problems across agent sessions.
• –Flagged production sessions can become evaluation scenarios without manually recreating test data.
• –Simulations run candidate agents in isolated sandboxes using the same tools and configurations.
• –Tool-call traces help distinguish reasoning failures from incorrect tool behavior.
• –Support is currently narrower for third-party frameworks, and Signals processing can take up to an hour.