Cohere's North 2 Wants to Run Your Agents, Any Model You Like

Cohere has released North 2, a substantial rework of the enterprise agent platform it has been quietly deploying for the past year. Announced on 5 October and covered by SiliconANGLE and HPCwire, the update is aimed less at dazzling demos and more at the unglamorous machinery that decides whether an agent survives in a regulated company: who can run it, what it is allowed to touch, and how much it is allowed to cost.

The headline change is orchestration. North 2 lets agents carry out multi-step tasks on their own and pull in a human for the steps that need a signature. Teams can build reusable agents and skills from plain prompts, then share them across an organization alongside common libraries of knowledge and assets. A new cross-session memory lets an agent keep context from one conversation to the next, rather than starting cold each time.

The controls are the point

What separates North 2 from the agent announcements crowding the calendar is the admin console. The North Admin tool tracks usage at the level of individual users and individual agents, sets consumption tiers and request-rate limits, and imposes organization-wide caps. Administrators can set alerts before a limit is reached and assign different models to different teams. On security, the platform applies guardrails per agent and scans both prompts and responses for sensitive data and attempted prompt injection.

That emphasis is deliberate. Cohere has spent the year putting North into finance, healthcare, telecommunications, manufacturing, energy and the public sector, and it says those deployments shaped the new release. The platform is model-agnostic, so a bank can run Cohere's own models or ones it already operates, and it can be installed in the cloud, on premises, or fully air-gapped for customers that cannot let data leave the building.

A control plane, not another agent

The positioning matters. Much of the industry is racing to ship the agent itself, from Google's universal Gemini worker to the proposals for common standards that would let agents from different vendors talk to one another. Cohere is betting that the harder enterprise problem is governing a fleet of them once they are loose inside a company, a concern that has pushed others, including Nvidia, toward sandboxing and oversight tooling.

Two caveats are worth keeping in view. Cohere has not published pricing, which is set per deployment, and it has released no performance benchmarks for the new version. For a product that asks to sit at the center of an enterprise's agent strategy, those are the numbers buyers will want before they commit.