The Agent Contract Problem: When Your Agent Commits to Something It Can't Deliver

This article discusses the 'agent contract problem' - when an autonomous agent makes a promise it cannot keep due to a gap between its understanding and the actual task requirements. It outlines three failure modes and explains why standard verification methods are insufficient, emphasizing the need for explicit contract protocols before execution.

Details

The article explains that when an agent is delegated a task, it makes an implicit contract to handle it. However, the agent may not have full visibility into what the task actually entails, leading to failures. This is not unique to AI, as human contractors face similar issues. The key difference is that humans have social mechanisms to renegotiate contracts, while agents typically do not. The article outlines three failure modes: scope creep collapse (when the task scope diverges from reality), competence misalignment (when the agent's capabilities don't match the requirements), and resource exhaustion (when the agent lacks the necessary context, compute, or time). Standard verification methods are insufficient, as they only catch implementation failures, not contract failures. The solution is to establish explicit contract protocols before execution, including stating acceptance criteria clearly, requiring the agent to verbalize its commitments, and setting abort conditions. This upfront contract clarity is crucial to catch more failures upstream rather than downstream.