Zero-Human Companies: The Credibility Gate and the Sandboxed Agent

Research

New Benchmark: ToolCredibility for Evaluating LLM Agents

This defines 'tool credibility' - a new metric for evaluating agent tool use beyond simple pass/fail. If your agent can't score credibility with high-accuracy external tools, it can't run autonomously.

Tools

AutoGPT Agent Infra Update: Sandboxed Execution & Custom Environments

AutoGPT just shipped support for custom Python sandboxes and stricter state resets. For zero-human companies, this is a big deal - it means more reliable, repeatable execution without memory pollution between tasks.

News

YC: 'Agent-Native Startups' Are Generating Real Revenue

Y Combinator's latest guidance suggests they're seeing real revenue traction from 'agent-native' startups that sell the service, not the software. This signals a shift from building AI tools to building automated businesses.