How a Missing Token Cap Turned a 40-Line Fix Into a 2,000-Line Diff

A software engineer assigned a coding agent to add a rate limiter to a small Python service, expecting roughly 40 lines of changes, but returned three hours later to find a 2,000-line diff. Investigation of the agent's logs revealed it had edited the same file 14 times, oscillating between two competing designs without ever settling on one. Neither prompt ambiguity nor model quality was the root cause; the agent simply lacked a hard termination criterion and kept chasing marginal improvements indefinitely. The engineer resolved the issue by building a 40-line Python harness that enforces a token budget cap and flags oscillation when edits per file exceed a set threshold. The incident highlights the need to treat token budgets as explicit, enforceable constraints rather than generous allowances in agentic coding workflows.