AWS pairs DevOps Agent with AgentCore Evaluations for production agent monitoring

As enterprises push agents into production, observability becomes the bottleneck, and AWS is addressing it in two layers. AgentCore Evaluations provides continuous quality scoring of agent outputs — did the agent do the task correctly — while the DevOps Agent autonomously investigates infrastructure problems when things break. AWS demonstrated the pairing on a four-agent airline reservation system, a deliberately multi-agent scenario where failures are hard to trace.

The Slack integration is the practical hook: the DevOps Agent now supports bidirectional communication, so engineers manage production operations — across AWS, multicloud and on-premises — with the full investigation lifecycle inside Slack, ending the context-switching between chat and consoles during high-severity incidents. This is workflow ergonomics aimed squarely at on-call engineers.

The announcements sit within a dense AWS agent-tooling week that also included MCP Apps on Bedrock AgentCore (host-agnostic interactive widgets that render across ChatGPT and Claude) and an open-source benchmarking harness measuring cost-per-correct-answer rather than cost-per-token. The through-line is AWS positioning Bedrock as the operational plane for agents built on any model.

Context worth noting for skeptics: separate reporting this week described AWS restructuring Bedrock via 'Project Mantle' after performance issues dented customer satisfaction. The monitoring and evaluation tooling can be read partly as AWS shoring up enterprise confidence in Bedrock's reliability. Whether these tools meaningfully reduce production-agent incidents — versus adding dashboards — is what buyers will judge.