Production agents require evidence loops instead of green dashboards

Datadog’s 2026 State of AI Engineering survey of 650 enterprise technology leaders found that 78% have at least one agent pilot running, yet only 14% have successfully scaled an agent to organization-wide use (Douglas Levin from Lessons from a Startup Life). Engineering teams often mistake a successful HTTP 200 response or a lack of tool errors for a completed task, even when the agent failed to check business constraints like seat availability or dietary preferences (Kevin Payne from Growth Guides). Anthropic recommends that teams assess the final state of external systems rather than requiring a particular sequence of actions to verify that an agent accomplished its objective (Tech Scoop).

Agus Sudjianto notes that an agent may recommend a credit approval correctly while simultaneously pulling data from an unauthorized source or skipping a mandatory verification step (Agus Sudjianto from Agus’s Substack). A production system requires an evidence loop outside the conversation that produced the final answer, where the agent generates a receipt containing the outcome, evidence, and an independent check (Kevin Payne from Growth Guides). This allows teams to distinguish between a model failure and a control failure by inspecting the execution chain, including retrieved evidence and enforced permissions, rather than relying on a final dashboard score.

Defining success before deployment enables teams to detect bad outcomes, as demonstrated by the need to specify whether a task is complete when a document exists in a folder or when it reaches an approved recipient (Douglas Levin from Lessons from a Startup Life). Teams should record decision logs including options, probabilities, and model versions to provide a concrete answer to why an agent performed an automated action (Bhavishya Pandit from WTF In Tech). "The useful idea is that production work needs an evidence loop outside the conversation that produced it."