Fetching from the wire…
Agents2026-09-19 · source-backed
arXiv 2609.19124 builds a deliberately small testbed: a hidden country flag is ground truth, each agent sees only a private crop, agents exchange beliefs and weigh peer evidence. Even at that scale it reproduces non-monotonic scaling with population size, gains from social-awareness prompting and team diversity, and strong organizational-structure effects. Small populations collapse onto one belief; larger ones polarize, and that polarization drives the large-population performance decline while creating belief diversity. arXiv They introduce social circuit attribution to predict which agent and which view drives the outcome, verified by causal patching.
Each link below shares sources, entities, or timing with this story.
A training-free multi-agent framework coordinating curriculum, actor and verifier agents to explore an unfamiliar environment and retain what it learns, including reusable causal relationships between actions, conditions and consequences (arXiv 2609.15364). Broad parallel self...
arXiv 2609.17598 studies PRs from OpenAI Codex, Devin, GitHub Copilot, Cursor and Claude Code across 2,807 repositories (Dec 2024 to Jul 2025), combining AIDev with 58,792 cached GitHub API responses. Codex PRs were reverted 6.1% of the time against a human baseline of 11.5% (...
The argument is that a static expert-written harness is the wrong shape, because one strict enough for a given model over-blocks a different one. It searches natural-language policies and executable code logic jointly against behavior analysis, domain specs and adversarial rev...
arXiv 2609.09560 ran 30 professional developers and advanced students through equivalent tasks under traditional, AI-assisted and AI-led conversational conditions with repeated-measures ANOVA plus thematic analysis. AI-led cut completion 27% against traditional and 12% against...
A case study tracked a repository catalog from a three-day agent-built hackathon prototype through public deployment (arXiv 2609.04711). Implementation was fast; making it trustworthy was not. The consequential problems were not crashes but plausible-but-wrong output traced to...
After 20+ years maintaining Paint.NET, Rick Brewster concluded WINE's Direct2D would never be complete enough for what he needed, so the app now carries its own from-scratch reverse-engineered Direct2D implementation. He puts it at 180,000 lines against 700,000 for the rest of...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.