Runtime incentives increase hard-task agent success from 0.60 to 1.00

Researchers published a paper on personal AI agents that operate for a long time and hold accounts, mailboxes and memory. They say these agents have no incentive layer. This is a rule in the control software that sets the model, the thinking time, the memory and the tool access of an agent. In a test with two models, the rule increased success on the hardest tasks from 0.60 to 1.00.

How it works

An incentive here is not text in a prompt. It is a contract in the control software, and the agent must use it. The contract sets the model tier, the thinking time, the memory, the tool access and the checks.

The numbers

• The test had 240 runs and 60 tasks that the authors can check.
• Prompt text about incentives had no detectable effect.
• Runtime allocation increased success from 0.60 to 1.00, at 4.8 times the price.

What is not known

The authors tested only two models. A simulation, not use at work, shows less gaming by agents. They say live deployment is future work. They released the code, the simulator and all results.

Insight: Control moves from the prompt to the runtime

The paper finds that words in a prompt did nothing, but resource rules in the runtime worked. Meta and Sierra also make an open protocol in which companies set limits on the work of an agent. This reads as a move of agent control from words to rules in the runtime, and the party that owns the runtime holds the control. Leaders must judge agents by the controls that they give.