AI agents are generating a new class of production incidents that existing incident management frameworks cannot classify or track. These failures occur when an agent executes an action that is technically correct within its given context, but that context is incomplete—creating cascading infrastructure failures that engineering teams don't recognise as agent-related problems. The issue isn't that agents are malfunctioning; it's that they're operating as designed within partial information sets, then triggering downstream consequences that traditional postmortem templates treat as infrastructure failures rather than agent decision failures. This represents a fundamental gap in observability: teams deploying Zendesk's autonomous service automation or similar agentic systems lack the instrumentation to trace failures back to incomplete context rather than technical bugs.
For CX teams, this creates an immediate operational blind spot. When an AI agent handles a customer interaction and makes a contextually reasonable but ultimately incorrect decision—such as issuing a refund, escalating to the wrong department, or triggering a workflow based on incomplete customer data—the failure gets logged as a support ticket, a process error, or a system glitch rather than an agent context failure. This means your team is likely already experiencing these incidents without recognising them as a distinct category. The question becomes: how many of your current ticket backlogs, escalations, and customer complaints are actually symptoms of agents operating on incomplete information rather than genuine support failures?
The implications extend beyond incident classification. As enterprises scale agentic AI across contact centres and multichannel operations, the absence of context-failure tracking creates compounding risk. Teams need to establish new observability practices that capture not just what an agent did, but what information was available to it when the decision was made. Without this, you're essentially running production systems where a significant failure mode remains invisible to your incident response processes—and your postmortem culture has no framework to address it.
There is a category of production incident that engineering teams are not tracking yet — because it doesn't fit any existing postmortem template. The agent initiated an action. The action was technically correct given the agent's context. The context was incomplete. The infrastructure casc