Greg Benson is chief scientist at SnapLogic and a computer science professor at the University of San Francisco who applies programming-language theory and distributed-systems research to enterprise AI. His Agent Continuations enable autonomous systems to pause for human approval, preserve their execution state, and resume complex workflows without repeating completed work.
Benson joined the University of San Francisco faculty in 1998 and earned his doctorate in computer science from the University of California, Davis, in 1999. His academic research encompassed operating systems, concurrent programming languages, and reliable distributed computation. He led development of USFMPI, a multithreaded Message Passing Interface implementation for Linux, and co-created FlashMob Computing, which assembled approximately 700 volunteer computers into an experimental high-performance computing cluster. He also contributed to River, a Python framework for prototyping parallel runtime systems.
Benson joined SnapLogic in 2010, helping develop its cluster processing, big-data infrastructure, cloud architecture, and machine-learning capabilities while continuing his academic work. His interests subsequently expanded into practical program synthesis and generative AI for enterprise integration. In 2023, he coauthored an Amazon Bedrock text-to-pipeline application that translated natural-language business requirements into executable integration workflows.
Making AI agents interruptible and reliable
- Saved computation for autonomous workflows. Agent Continuations adapt a programming-language concept to AI execution: preserve an agent’s message history alongside metadata identifying its pending action, suspension condition, and precise restart point. A workflow can then resume from its captured state instead of starting over.
- Human-in-the-loop approval before consequential actions. Agents suspend before sensitive operations such as granting account privileges, transferring money, or deleting accounts. In Benson’s employee-onboarding example, a subagent creates an account but waits for human authorization before assigning privileges.
- Resumable multi-agent workflows without idle infrastructure. Continuation objects recursively preserve orchestrators, subagents, message histories, and pending tool calls. Approval requests propagate to the application layer, while execution loops shut down completely until a response arrives. The same architecture supports checkpointing, failure recovery, rate-limit interruptions, and application-defined suspension conditions.
Benson and SnapLogic’s research team implemented these ideas in AgentCreator and a public Agent Continuations prototype built on the OpenAI Python API. His technical explanation of resumable workflows and AI Engineer World’s Fair presentation establish the practical objective: autonomous agents that remain recoverable, resource-efficient, and subject to meaningful human control.
