OpenAI published formal Lean 4 proofs for ten long-standing open problems in group theory, cryptography, and quantum complexity, each solved by an internal Astra instance for under $2,000, with machine-checkable certificates on GitHub.
Why it matters: Lean-verified outputs set a new credibility bar for AI-generated work — enterprise buyers evaluating AI for high-stakes tasks will increasingly expect formal verification, not just benchmark scores.
Microsoft's agentic security platform goes into public preview today: red agents map attack paths, blue agents triage risk, green agents remediate — running on MAI-Cyber-1-Flash, which scores 96% on CyberGym at half the cost of current configs.
Why it matters: The first major-vendor production multi-agent system to reach GA this week. Its specialized-agent-plus-in-house-model architecture is the pattern practitioners will use as a reference when scoping their own agentic builds.
MIT-licensed and cloud-first with native Slack and web UI, cross-role scoped memory, and support for Claude Code, Codex, OpenCode, and Pi — YC uses it across accounting, legal, events, and engineering.
Why it matters: A production-proven, company-wide agent harness released under MIT is the kind of reference architecture operators ask for when evaluating whether to build their own agent stack rather than buy a vendor solution.