Stopping a runaway refactoring agent loop that never converges

CCDV-F Agents & Workflows Medium Question 36 of 105

An autonomous refactoring agent runs shell commands and edits files in a loop. On a repository with circular test failures it re-runs the same two commands for hours without converging.

Which control stops the runaway loop?

Show answer and explanation

DCap the loop at 40 iterations plus a token budget, and halt on a repeated tool call

Stopping conditions such as a maximum number of iterations are what hold an agent loop, and a token budget plus a check for a repeated tool signature catches this circular case specifically. A per-request max_tokens bounds one response, not the number of responses, and more container memory lets the loop run longer. A prompt rule asks the model that cannot tell it is looping to notice that it is. tool_choice auto is already the default and leaves termination to that same judgement.

Source: anthropic.com/research/building-effective-agents

Anonymous, no account needed. Your tally stays in this browser.

Next question Start 5 timed questions Browse all 105 original CCDV-F practice questions