Type: url
Original: https://cposkitt.github.io/files/publications/agentspec_llm_enforcement_icse26.pdf
Status: generated
Summary
This ICSE 2026 paper by Haoyu Wang, Christopher M. Poskitt and Jun Sun (Singapore Management University) introduces AgentSpec, a lightweight domain-specific language and framework for specifying and enforcing runtime safety constraints on LLM agents through rules made of a trigger, predicate checks and enforcement actions. Implemented on LangChain and evaluated on code-execution, embodied and autonomous-driving agents, it is reported to prevent unsafe executions in over 90% of code-agent cases, eliminate hazardous actions in embodied-agent tasks and enforce full law compliance in the tested driving scenarios with millisecond-level overhead. The paper also evaluates rules generated automatically by OpenAI o1.