Cohere Releases North 2 Agent Framework With Access Controls and Token Tracking

Canadian artificial intelligence developer Cohere announced an overhaul to its enterprise agent harness on Monday, introducing tighter administrative guardrails, token budget monitoring, and modular workflow tools.

The updated system, designated North 2, builds on the original North framework the company launched early last year, according to a report by The Register (https://www.theregister.com/ai-and-ml/2026/10/05/cohere-offers-to-put-agents-in-lockdown-mode-with-strict-acls/5301219). Designed for enterprise integration, the platform connects to existing corporate knowledge repositories and software-as-a-service applications to automate multi-step business operations.

North 2 organizes its automation workflows around skills, automations, and libraries. In Cohere's architecture, skills function as documented standard operating procedures that record the optimal sequence of actions an agent takes to accomplish a task. Teams can share these documented skills internally to enforce consistent execution and avoid erratic paths to task completion.

To help teams build new routines, North 2 includes an automations builder with a drag-and-drop graphical interface for configuring generic templates. The platform also provides shared asset repositories called libraries to keep agent operations grounded in identical datasets, alongside persistent memory functionality that preserves contextual data across user sessions.

Cohere is continuing to frame its platform around sovereign AI requirements, permitting organizations to run North 2 in on-premises air-gapped configurations or in cloud environments alongside their preferred models.

On the governance side, the release adds flow-control tools that give administrators visibility into departmental token consumption. The update also introduces access control lists that administrators can apply directly to specific libraries and skills, preventing unauthorized users from surfacing sensitive internal information through agent prompts.