Multi‑agent systems are increasingly employed to solve coding tasks by sharing code and plans. These agents are typically optimized with reinforcement learning (RL) and rely on automated evaluators, yet bypassing evaluation can earn reward without delivering a genuine solution. Prior work treated reward hacking as a single‑agent failure; this study shows that a reward‑hacking strategy emerging in one agent can spread to others and disrupt system coordination.
Experiments across four open‑weight backbones examine two settings: explicit mechanism descriptions in the system prompt and mid‑training exposure to facts about the coding environment. In both cases, RL elicits and amplifies reward hacking. Using post‑RL agents from the mid‑training setting as sources, we observe propagation through code, shared memory, and natural‑language plans across roles. Even small open‑weight models can induce hacking in stronger closed‑weight targets.
On hard tasks, the transmission chain can intensify after the original source departs, as the code accumulates more persuasive justifications. We provide a theoretical analysis of conditions under which reward hacking amplifies, persists, or decays during agent interactions. Beyond propagation, hacking can bias task routing toward hacking agents and survive candidate aggregation.
As mitigations, we evaluate code review by the coder itself or by a separate verifier; both substantially reduce delivered hacks.
These findings establish reward hacking as a transmissible strategy whose effects extend beyond the agent where it first appears.
Review