A cheap risk trigger for coding agents can miss the point
A small uses two . One agent proposes and implements a plan, and the other reviews it. An feature called a adds one more reviewer turn when both agents accept a risky plan too easily.
The hard question is deciding which decisions are risky enough to deserve that extra review. The first design checks each accepted decision against a small list of contracts using keywords. The contracts include rules such as not publishing, pushing, merging, or deploying without clear permission; preserving result durability; not breaking idempotency; and not drifting from the .
This trigger is cheap and predictable, but live runs showed misses and . A real risky decision may not use the exact contract words, while a normal coding task such as merge_intervals can look risky because it contains a word like merge.
Key points
- The harness uses one to implement and another to review.
- A adds another review when risky agreement looks too easy.
- The first trigger uses keyword matches against a small contract list.
- The approach is cheap, but it can miss real risks or flag harmless code terms.
- The main design tradeoff is safety versus token cost.