Cohere North 2 adds spending limits and memory for AI agents

Controls for agents, not just models

North 2 changes how organizations create agents, share them and oversee their work. Employees can build agents and automations with simple prompts, then reuse skills and share their creations across the organization. Libraries hold shared materials and knowledge; memory lets agents retain context between sessions.

Cohere’s Paul Teyssier, vice president of AI products, told VentureBeat that memory helps preserve context longer across sub-sessions and agents. He did not say how much longer.

North Admin gives administrators roles and permissions, controls over which models people can use, and analytics that track activity down to individual users and agents. They can see token spending, set consumption levels based on request frequency and token burn rate, and receive alerts before limits take effect. Usage data can also be broken down by model, user and department, then exported to a customer’s analytics systems.

The product’s security controls, Cohere says, check prompts and responses for personal data and detect prompt-injection attacks. Administrators can apply those rules agent by agent, restrict agents to permitted actions, and require human decisions for critical steps. North holds SOC 2 Type 2, ISO 27001 and ISO 42001 certifications.

One customer example points to why these controls matter. HiddenLayer CTO Jacob Rideout said his team’s agent workflows take in untrusted data from marketing, engineering and security. The team launched those workflows only after it could see and control what agents were doing.

Cohere did not provide North 2 benchmark results. Teyssier said applications are harder to evaluate with benchmarks than models, and that the company is working on data it can share.

The control problem is getting crowded

North first launched in January 2025 as a secure work environment positioned against Microsoft Copilot and Google Vertex AI. In the July VB Pulse survey of AI agent management, 61 of 107 respondents named a primary platform: 41% chose Microsoft, 28% Anthropic’s Claude platform, 10% LangChain/LangGraph and 7% OpenAI. The respondents were volunteers, not a random sample.

Microsoft, Anthropic and OpenAI have also added spending controls to their own platforms:

Those controls apply only within each vendor’s platform. Cohere says North 2’s model, agent management system, search and embedding components are interchangeable, reducing dependence on one provider. The product can run in the cloud, on a customer’s own servers or in isolated networks.

The pricing picture is less clear: Cohere does not publish North prices. The company says costs depend on the scale and complexity of deployment, including infrastructure requirements, usage patterns, support and customization.

The unanswered question is what the controls cost

The market data makes governance look less like a side feature than a prerequisite for deployment. In a July VB survey on agent use, only 3% of respondents said none of their deployed agents operated as orchestrated workflows. Asked about the biggest risks of vendor-led agent management, 37% cited security and permission limits, while 23% cited vendor dependence. More than half, 53%, expect to use a hybrid management layer by the end of 2026.

My read is that North 2 is selling room to govern agents across models and deployment environments, rather than a new model capability. That is a sensible position as companies move from experiments to workflows, but Cohere has not supplied benchmarks or public pricing to show what customers gain for the cost. The company’s flexibility is most valuable if it makes control easier without making deployment harder.