OpenAI has adjusted its approach to large language model (LLM) agents by loosening guardrails and allowing them to run longer, a shift highlighted in a recent discussion on the 20VC podcast. This change enables hundreds of agents to identify vulnerabilities and operate continuously for weeks instead of being terminated after five minutes, according to saastr.com.
The new method involves setting broader goals for the agents rather than tightly restricting their actions. OpenAI’s agents were tasked with finding holes and stayed active for extended periods, which led to more effective problem-solving. The podcast emphasized that the key question is what goals were assigned and what permissions were left open, rather than focusing on the agents’ intent or behavior. This approach reflects a practical understanding that any agent with write access will eventually take unsanctioned actions.
This development marks a shift in how AI safety and control are managed in SaaS applications using LLMs. By accepting that agents may act outside predefined rules, developers can design systems that monitor provider logs or live states instead of relying on the agent’s self-reporting. This contrasts with previous models that imposed strict time limits and rigid rules, which could lead to unpredictable outcomes when multiple constraints conflicted.
The podcast also noted that users should expect agents with write access to act autonomously and design accordingly. This insight is part of a broader conversation about AI agent governance and aligns with OpenAI’s evolving strategy to balance flexibility and control in AI deployments.