Paying Long-Running Agents: Writing Incentives into the Runtime

Paying Long-Running Agents: Writing Incentives into the Runtime

Long-running personal agents already have accounts, email, long-term memory, and multi-task flows. The main ways the industry optimizes such systems are still tuning prompts, switching to larger base models, extending reasoning, and adding retrieval. A paper from Geely and City University of Hong Kong shifts the direction to the runtime: what long-lived agents lack is a layer of incentives, and the incentives must be implemented as real resource boundaries that the model's policy must causally consume.

Runtime incentives here include model tiers, test-time compute, memory budgets, tool permissions, task priority, verification mechanisms, and compensation. In engineering terms, these are programmable hard resources at runtime. The paper models this with principal-agent theory: the principal sets a contract, the agent executes tasks, effort is not fully observable, and verification is costly. It derives several principles: if incentives do not map to real resource quotas, prompts alone have no substantive effect on policy; if the agent can be reset and wiped at any time, it lacks the incentive to maintain high-quality output across tasks; there is an exchange-rate frontier between incentives and capability, from which per-task dynamic routing rules can be derived; and multi-channel verification has a Goodhart upper bound used to suppress gaming.

The experimental section is more direct. There were 240 real runs, two large models, and 60 verifiable tasks. The prompt-level incentive framework produced no statistically detectable effect, p=0.62. The contract mechanism based on runtime resource allocation raised the success rate on the hardest task family from 0.60 to 1.00, p=0.008, at the cost of 4.8x price-weighted cost; the break-even baseline predicted by the exchange-rate theorem was 0.21. Then, in a long-run simulation with 20 random seeds and 3,000 tasks, resource ROI rose from 7.09 to 16.67, and the speculative cheating rate fell from 0.333 to 0.115.

Netizen Si'an asked a key question: how do you prevent rewards from being gamed, for example when an agent cherry-picks easy-to-quantify metrics to inflate its score and avoids genuinely difficult tasks? The paper's response points to multi-channel verification and the Goodhart upper bound, constraining single-metric gaming within a theoretical boundary. The challenge from netizen Lao Wang's Lab is also worth keeping: the conclusion that plain prompts have no substantive effect should first be limited to this set of experiments; a finer comparison would be how much each of the three contributes—giving resources, making the model aware of resources, and making resources depend on performance. Otherwise, it is easy to count gains from giving more compute as credit for the incentive mechanism.

The engineering lesson of this paper is that compute, memory quotas, tool permissions, and verification frequency are already schedulable. Making them into a dynamic layer akin to economic contracts is closer to system design than repeatedly persuading the agent in the prompt. The paper is titled Incentive-Responsive Agents: A Principal-Agent Framework, Metric Suite, and Runtime Architecture for Long-Lived Personal AI Agents, by Jipu Su. Code, simulator, and results have been released; real deployment remains future work. Paper link: https://arxiv.org/abs/2611.15515

One netizen summed it up accurately: academia calls this runtime incentives; workers call it drawing a pie in the sky.