A broad upgrade to Cohere’s agent platform
Cohere has introduced North 2 as a major update to its enterprise agent platform. Announced on 5 October, the release combines a redesigned orchestration system with shared knowledge, reusable skills, memory, application creation and automation. Cohere says the product reflects experience deploying North in finance, healthcare, telecommunications, manufacturing, energy and the public sector, where security and governance are part of the workflow rather than optional add-ons.
The platform is intended to let employees create agents and automations using natural language while keeping human oversight for critical decisions. Reusable components can reduce duplicated work across departments, and a visual workflow builder can make process design accessible beyond software teams. The practical test is whether these abstractions remain understandable when an automation fails, uses the wrong source or reaches a step that requires an exception.
The new harness coordinates multi-step work
At the centre of North 2 is an agent orchestration system for more complex, multi-step processes. Skills package capabilities, libraries hold shared knowledge and assets, and memory carries context between sessions. Applications can produce documents, dashboards and lightweight tools, while automations include templates, monitoring and connectors. This brings several parts of enterprise agent development into one managed environment.
Combining those parts can simplify deployment, but it also concentrates control. Administrators need version history for skills, prompts and automations, along with a record of which version ran. Shared memory should be scoped so one team cannot expose information to another. Connectors to Slack, SharePoint, Outlook, Jira, GitHub and other systems need least-privilege permissions and reliable revocation when a user or agent changes role.
Model choice is a deliberate feature
Cohere describes North 2 as model-agnostic: customers can use Cohere models or bring alternatives that better suit a task. Model choice can help an organisation balance quality, latency, cost and data requirements instead of forcing every workflow through one provider. It may also reduce migration risk when a model is retired or a better option becomes available.
Portability is not automatic. Different models interpret tools, structured output and long context differently, and evaluation results may not transfer. A workflow should be tested against every approved model configuration, with clear fallback behaviour. The platform also needs to expose which model handled each step so users can reproduce an output and cost reports can show where consumption occurred.
Deployment and autonomy controls address regulated use
North 2 supports self-hosted, private-cloud, hybrid, on-premises and air-gapped options. Cohere highlights security certifications, response guardrails and autonomy policies that restrict which actions an agent may take. Flexible placement is important for organisations with residency or confidentiality obligations, although private deployment still requires disciplined patching, monitoring and incident response.
Autonomy policies should be enforced outside the model's own instructions. A model can misunderstand a rule, while a tool gateway can deny an unauthorised action. Critical steps need explicit approval and a visible explanation of what will change. Air-gapped deployments also require a secure process for importing model updates and vulnerability fixes. Control over location does not remove the need for control over change.
Cost governance becomes a product feature
Governance should extend to templates shared across the organisation. A reusable skill can spread a useful method quickly, but it can also spread a faulty instruction or excessive permission. Owners should approve high-impact components, test changes against representative cases and retire obsolete versions. North 2 will be safer when reuse includes provenance and review, not only convenience.
Administrators should connect cost controls to business ownership. A token cap that stops an essential workflow without warning can be as disruptive as uncontrolled spending, while a generous shared pool may hide one poorly designed agent. North 2's analytics should let teams allocate usage to a department, application and outcome, with alerts before a hard limit is reached. Chargeback or showback can encourage better design when it includes the cost of retrieval, tools and infrastructure as well as model tokens. Teams should also test what happens at the limit: the agent needs to save state, report incomplete work and avoid leaving a partially executed business process.
North Admin adds token-rate limits, user and group quotas, organisation-wide caps, permissions and real-time analytics. These controls respond to a genuine problem: an agent can generate many requests through loops, parallel workers or repeated retrieval. Budget limits should be paired with alerts and task-level reporting so teams can distinguish valuable heavy use from an automation that is stuck.
North 2 is an ambitious effort to package intelligence, orchestration, security and financial control into one enterprise platform. Buyers should test complete workflows, including failures, model swaps and permission changes, rather than judging the system by a successful demonstration. Cohere has made governance and deployment ownership central to the product story. Measured reliability, transparent costs and auditable actions will decide whether the promised reduction in enterprise compromise is realised.