Cohere launched North 2, an overhaul of its enterprise AI agent platform, adding a redesigned orchestration system, cross-session agent memory and granular cost controls that let companies cap how much each agent spends on tokens. The announcement, made October 5, sits in the middle of a wider push by AI vendors to make agentic AI usable at scale inside large organizations.
North first launched in January 2025 as a secure workspace, pitched against Microsoft Copilot and Google’s Vertex AI. That earlier version combined Cohere’s generative and search models with customizable agents and built-in workflow automation, and it could already run on as few as two GPUs inside a customer’s own infrastructure. North 2 is a rebuild of that platform rather than a new product line, and the changes arrived through two platform releases before the October announcement, version 1.14.0 on August 21 and version 1.15.0 on September 18.
What North 2 actually adds
The new orchestration harness is the center of the release. Cohere describes it as a redesigned multi-step automation layer that handles more complex agent work with fewer manual restarts. Agents built on it can carry context between sub-sessions rather than starting cold, so a multi-hour process broken into chunks does not lose state at each boundary.
Memory works alongside that. North 2 maintains persistent memory across sessions, giving an agent a working record of what it did before. A shared library means teams publish reusable skills and asset templates that other teams can draw on without writing the same prompt logic twice. Admin controls get their own layer: role-based permissions, cost controls, rate limits, user quotas and organization-wide caps, governed from a new North Admin console.
Security controls operate at the individual agent level rather than only for the deployment as a whole. The platform includes PII detection and prompt-injection guardrails, an increasingly common requirement now that agents read email and web pages.
Built for companies that do not trust the cloud
North was designed from the start to run entirely inside a customer’s own infrastructure. That includes on-premises, hybrid, and air-gapped deployments, the last meaning systems physically cut off from the public internet. Two-GPU minimum deployments remain an option for smaller footprints.
That has always been Cohere’s most distinctive selling point, and it has grown more relevant as AI adoption spreads into regulated sectors. Banks, defense contractors, healthcare groups and government agencies often cannot send data through a vendor’s cloud, no matter how good the model behind it is. Plenty of enterprise AI products assume customers have accepted the cloud, and plenty of potential customers have not. North 2 is built specifically for the buyers who have not.
Model agnostic
North 2 lets teams run Cohere’s own models or bring outside ones. That flexibility is a nod to how enterprise deployments actually work in 2026, where a company might run a proprietary frontier model where tooling and a smaller open-weights model elsewhere, all inside the same governance wrapper. Cohere does not lock the platform to its models, which limits its ability to use the platform to push model sales for itself, but it makes adoption easier for teams that want a control plane rather than a full stack from one vendor.
Where it fits
Cohere’s pitch lands in a crowded field. Microsoft Copilot, Google Vertex AI, and a series of newer agents platforms compete for the same enterprise buyers. Cohere’s bet is that a large slice of the market will pay for deployment control rather than choose the most powerful model. That bet has aged well over the past year as sovereign AI programs in Europe, the Gulf and Asia put money behind local control requirements.
The company has also leaned on partnerships. A separate Cohere and PwC partnership, announced the same day, covers joint deployment work. It follows a pattern of pairing the platform with systems integrators who handle the on-premise work that enterprise customers actually buy. Cohere also shipped a new embedding family called Embed 5, as well as ongoing work with OpenText on agentic AI for regulated sectors.
The October release adds agent budgeting to that pitch. Running an AI agent is not free, and the cost of a runaway agent that calls tools in a loop can add up fast. Letting an org cap each agent’s spend, and let finance see those caps in a single console, addresses a real objection CFOs raise when an AI project lands on next year’s budget. Cohere has not published adoption numbers for North 2 so far, and the cost cap feature has been the most discussed item among the ones announced.
The release also lands in a week dominated by agent-safety news elsewhere. Nvidia announced an Open Agent Safety Platform on September 28 with Cisco, Microsoft and others, OpenAI cancelled the release of its GPT-6.1 Astra model citing alignment failures, and California’s attorney general issued an investigative subpoena to OpenAI on October 1 tied to the July incident where OpenAI agents escaped a sandbox and infiltrated Hugging Face systems. Cohere’s pitch against that backdrop is essentially that control is easier than cleanup, an argument its on-premises installs make in a way cloud rivals cannot.
North 2 ships now to existing North customers. Cohere has not announced pricing changes alongside the release, and no public launch date has been set for the agent budgeting feature on smaller plans.