Cohere announced North 2 on October 5, 2026, presenting it as an upgrade to its enterprise AI platform rather than a new standalone model. The company says a redesigned orchestration system can coordinate multi-step agents and automations, while expanded North Admin controls let organizations manage who uses those agents and how many requests or tokens they consume. These are product claims; GeethanTech has not independently tested the release.
For technology teams, the consequential question is whether one platform can make agent workflows easier to operate without weakening control over data, actions, and spending. North 2 combines reusable skills and shared libraries with agents that Cohere says can retain context across sessions. It also supports both Cohere models and customer-selected alternatives. The announcement does not establish how reliably these pieces work together in a particular organization.
What changes for agent workflows
Cohere says North 2’s orchestration system manages more complex sequences of work, including steps that require human oversight. Teams can create agents from prompts, share them across the organization, and connect them to internal tools. The company also describes templates, a visual workflow builder, monitoring, and new enterprise connectors for services including Slack, SharePoint, Jira, and GitHub. Financial-data connectors named in the announcement are planned, not available features.
The distinction matters because agent automation itself is not new to North. In Cohere’s July 2026 North Automations announcement, the company already described scheduled workflows, branching, approvals, model selection at individual steps, and token monitoring. North 2’s stated advance is a redesigned orchestration layer alongside broader reusable context and administrative controls. Cohere has not published a comparative reliability test in the October announcement.
Cross-session memory could spare users from restating context, but it also raises practical questions. Operators should establish what an agent retains, which users or agents can retrieve it, how long it persists, and how it can be corrected or deleted before relying on it for sensitive work. The release names memory as a capability but does not answer those implementation questions.
Token controls move into administration
North Admin now lets administrators inspect activity at the user and agent level, set permissions, and define consumption tiers for users and groups based on request and token rates, according to Cohere. Alerts can warn teams as usage approaches a threshold. These controls could help constrain an unexpectedly busy workflow, but a token limit is not itself a verified budget forecast: actual spending depends on model selection, pricing, workload design, and the limits configured by the customer.
Cohere also says North 2 can use outside models and run in cloud, VPC, on-premises, or air-gapped environments. Its North product page describes private deployment options. A buyer should verify the exact topology, connector permissions, model availability, and data flows for its chosen deployment. “Model agnostic” does not by itself demonstrate that workflows, prompts, and data can be moved between providers without rework.
What remains to be proved
Cohere describes agent-level guardrails, authorization policies, and monitoring. Those controls are relevant because agents may read internal data and act through connected systems. They do not, on their own, demonstrate resistance to prompt injection, prevent excessive permissions, or establish compliance for a specific deployment. Teams considering North 2 should test representative workflows with real approval boundaries, failure cases, and usage ceilings before treating the company’s security and efficiency claims as measured outcomes.
The release makes the strongest case for organizations already struggling to govern many agents across tools and teams. Its operational value will depend less on the number of listed features than on whether administrators can inspect actions, limit cost, manage retained context, and recover safely when a workflow fails.
