Retry or escalate: rules for an agent that fails at SQL

On 2026-11-07, the AI section of DEV Community published an article titled "Retry the cheaper model, or upgrade?

Decision rules for agents after a SQL failure." The piece argues that when a cheaper model keeps emitting the same unknown column name, repeated retries only burn through budget, return nothing usable, and teach you nothing about the schema. In a drafting loop, a failure the model was never able to fix can consume the entire allowance: a cheap tier that repeatedly returns a column name absent from the catalog ends up, four repairs later, with no usable output, a spent budget, and no added knowledge of the schema. The article therefore proposes deciding by failure type whether another attempt is warranted on the same tier at all, rather than a blanket choice between retrying and upgrading. The author also references an earlier post about an agent-written query that locked a production table. Based on the reports received so far, that is the latest stage this topic has reached.