Error-Tolerant Agents

The use of agents in today’s Internet world is expanding rapidly. Yet, agent developers proceed largely under the optimistic assumption that agents will be error-free. Errors may arise in agents for numerous reasons — agents may share a workspace with other agents or humans and updates made by these other entities may cause an agent to face a situation that it was not explicitly programmed to deal with. Likewise, errors in coding agents may lead to inconsistent situations where it is unclear how the agent should act. In this paper, we define an agent execution model that allows agents to continue acting “reasonably” even when some errors of the above types occur. More importantly, in our framework, agents take “repair” actions automatically when confronted with such situations, but while taking such repair actions, they can often continue to engage in work and/or interactions with other agents that are unaffected by repairs.