Vals AI compared solo agents with lead-plus-subagent teams on 50 web-app builds using GPT-6 Sol and Claude Opus 5.5: teams cost 1.8–5.1× more, while only one of four comparisons showed a statistically significant quality gain. The study’s task-specific setup does not settle when parallelism helps, but it is a useful counterweight to default “agent swarm” architectures; most extra cost came from repeatedly sending subagent context. Results are independently evaluated by Vals, not a broad cross-task finding. Source: https://www.vals.ai/blogs/multi-agent-vibe-code-bench
Anthropic launched OSS Scanner, an opt-in, free service that periodically scans participating open-source repositories with its strongest models and sends vulnerability reports through maintainer-facing coordinated disclosure. Novel infrastructure angle: turning frontier-model vulnerability discovery into a recurring upstream-maintenance service; reports are not human-reviewed by default on the fast track, while Anthropic says its standard workflow includes human review. Sources: https://www.anthropic.com/research/launching-opt-in-vuln-finding-service-for-open-source ; https://news.ycombinator.com/item?id=50010549