Enterprise AI has a continuity problem. A company can buy access to a stronger model every few months and still rebuild the same instructions, integrations and working practices around it. The expensive knowledge is often not in the checkpoint. It is in how the organisation gets a useful result from its own systems.
Cohere’s North 2 launch on October 5 targets that surrounding layer. The company describes a redesigned agent harness with memory, reusable skills, shared libraries, automations and artifact creation. North is also model-agnostic. The interesting possibility is that a workflow can retain its institutional knowledge when the engine beneath it changes. The less obvious consequence is that the workflow becomes a lasting system with its own state and failure modes.
The release is an orchestration upgrade
North 2 is presented as a full-stack platform, not a single new language-model checkpoint. Cohere lists cross-session memory, reusable agents, shared knowledge, application creation and a visual workflow builder. Deployment options include private cloud, on-premises and air-gapped environments, while administrative controls include quotas and cost limits.
Those are company product claims. They do not establish that every model will behave identically inside the same workflow or that every installation automatically satisfies an organisation’s requirements. The distinction matters because platform continuity and model interchangeability are different promises. One can preserve a workflow definition while changing how reliably its steps are executed.
A reusable skill can become a shared dependency
Saving a useful procedure once can prevent employees from rebuilding it repeatedly. It can also spread a mistaken assumption. A skill created for one department may rely on a field name, approval rule or document format that another department interprets differently.
This hypothetical example illustrates a governance need created by reuse itself. A shared library should carry scope and ownership, not merely a helpful title. When a procedure changes, teams need to know which agents depend on it. The cost of maintaining that dependency graph is part of owning an agent platform, even when the underlying model remains a replaceable service.
Persistent memory needs a correction path
Cross-session context can save an agent from repeatedly asking the same questions. It can also carry yesterday’s answer into a situation where it no longer applies. Deleting a conversation may not be enough if its conclusions were compressed into a skill, a shared record or a workflow default.
A practical evaluation should therefore include a correction task. Give the system an accepted fact, let it influence a reusable workflow, then revise the fact and inspect later execution. This is a proposed acceptance test, not a North 2 benchmark result. It asks whether the platform can update the consequences of retained knowledge, rather than only show that it remembers.
Model choice becomes one layer of ownership
Cohere’s emphasis on deployment control and a model-agnostic platform is significant because the organisation’s learned practices need not be tied to one checkpoint. But portability should be tested: replace the model, replay representative workflows and inspect tool arguments, decisions and final artifacts. A procedure written for one model may contain assumptions another model does not follow.
The release also distinguishes shipped connectors from planned financial-data integrations; those should not be merged into one availability claim. The larger buying question is how the platform preserves useful organisational state while keeping it inspectable and changeable. An enterprise does not gain control merely by owning where inference runs. It also needs to know which memories and procedures now govern its agents. North 2 makes that surrounding layer central to the product, turning continuity into an asset that must be evaluated alongside intelligence.
Cohere official release and methodology
References
Questions, news tips or corrections? Contact [email protected].
