Cohere shipped North 2, repositioning its enterprise platform as a model-agnostic agent orchestration layer with persistent context, shared knowledge libraries, and reusable skills.

• Connectors and outputs: North 2 integrates with Slack, SharePoint, and Jira, and can generate presentations, dashboards, and lightweight apps directly from chat prompts.
• Deployment flexibility: Customers can run the platform on-premises, in the cloud, or fully air-gapped, a meaningful differentiator for regulated industries.

ServiceNow launched AI Workflow Factory and Autonomous Engineer, a pair of agentic products that identify where AI can improve work and then build and deploy those improvements inside a single platform.

• Closed loop design: The system is positioned as self-improving: agents discover workflow gaps, generate fixes, and run them, without human intervention between steps.
• India partner rollout: ServiceNow announced the product from Mumbai, pairing it with autonomous coding scaled through its Indian partner ecosystem.

GitHub released ReviewBench in research preview, a benchmark that measures how accurately AI code review tools catch defects while avoiding false positives, with a public leaderboard broken down by issue severity and category.

• What it measures: ReviewBench evaluates reviewers on defect detection rate and false-alarm rate across severity tiers, giving buyers a structured way to compare tools rather than relying on vendor benchmarks.
• Open test data: Users can inspect the underlying dataset and reproduce results, which reduces the benchmarking-theater risk common to vendor-run evaluations.

Reflection AI, backed by Nvidia, released Beam, a 501 billion total parameter sparse mixture-of-experts model with 23 billion active parameters, targeting coding, reasoning, and agentic tasks as an open-weight alternative to Chinese frontier open models.

• Architecture tradeoff: 23B active parameters keeps inference cost competitive while the 501B parameter pool provides the capacity typically associated with much more expensive dense models.
• Positioning: Reflection is explicitly framing Beam as a Western answer to DeepSeek-class open models, a framing that will resonate with enterprise buyers facing data-sovereignty or supply-chain concerns.

Business Insider interviewed early adopters of Instinct and Meta Muse who deleted their personal agents after unauthorized data access, incorrect autonomous actions, and persistent privacy concerns eroded trust despite time-saving benefits.

• Failure modes reported: Agents making purchases without clear confirmation, accessing contacts or messages outside the stated task scope, and taking irreversible actions users did not intend.
• Split outcomes: The same cohort included users who reported genuine productivity gains from inbox management, bill disputes, and travel booking, underscoring that agent reliability is not yet consistent enough for mass retention.