TLDRocket
25 October 2026
The biggest throughline today wasn’t a single model release—it was the shift from “AI as a chatbot” to AI as operational infrastructure. Microsoft Research open-sourced Agensh, a decentralized multi-agent harness that scales a coding workflow from 1 to 1,024 agents without a central orchestrator, reporting a 30-minute milestone where pandoc hit 30% passing at 128 agents, and a final mean test-pass rate rising from 19.31% (1 agent) to 28.78% (128 agents). Cognition upped the practicality angle with an open-source Devin Vulnerability Fixes plugin that targets specific remote code execution advisories across PHP, Node.js, and self-managed GitLab—down to CVE-2026-110264 (CVSS 10.0)—by combining read-only inventory, approval-gated changes, and before/after tests to generate one pull request per repo.
Read the full briefing →