
Architecting Trust Without Killing Autonomy: Transactional Envelopes and Microservice Resilience Patterns for Reliable Computer-Using Agents
Purushothaman S, Madhukar Irvathraya • Nov 14 2026
Computer-using agents can do impressive work, but they still break halfway through workflows and leave systems in a messy, half-finished state. This paper adds a transactional layer around the agent instead of locking it into rigid scripts. The layer borrows microservice patterns like Sagas, circuit breakers and idempotency so that every change the agent makes can be committed, rolled back or escalated to a human. The agent keeps its freedom to explore, and enterprises get the recoverability and audit trail they need.