New guardrails in North 2 monitor token usage by department and implement access control lists to restrict data access within libraries and skills.

Security and governance are reinforced by ongoing testing, red-teaming, third-party scans, observability tools, and collaboration with Bell Cyber to elevate government-grade security for agentic AI.

Cohere unveils North 2, an upgraded enterprise AI agent platform designed to improve task orchestration and enforce token spending controls.

North 2 is positioned against major enterprise AI players by emphasizing model portability, security controls, and reduced vendor lock-in through interchangeable components—model, harness, search, and embedding.

Beyond a chatbot, North 2 acts as a software layer with an agent harness that connects AI models to applications to complete tasks, using an engine-and-vehicle metaphor shared by Cohere’s CIO.

Shared libraries enable cross-type resource sharing and beta API access to libraries, centralizing resource management across teams.

Skills document task procedures, automations are drag-and-drop templates for workflows, and libraries ensure agents use common datasets to ensure consistent outcomes.

Security and control are prioritized for regulated industries and public sector clients, with features that restrict actions, require human input for critical tasks, and define access and usage parameters for agents.

Early adopters include Bell Canada’s security operations unit and LG CNS in South Korea, deploying North for internal use and customer-facing applications.

Industry context shows enterprises seek real-time spend controls and reduced context gaps in agent responses, with many pursuing hybrid control planes by 2026, driving demand for governance and cost control.

North 2 meets GDPR and SOC-2 standards and targets functions across HR, finance, and customer support.

North Admin introduces governance capabilities with roles, permissions, model-usage controls, real-time analytics, token-spend visibility, alert thresholds, and flow-control across consumption tiers; it remains model-agnostic and supports on-premises or air-gapped deployments.