Cohere launches North 2, giving business AI agents memory and spending caps

North is a software layer that sits around AI models and lets companies build and run agents on their own data. It has been generally available since August 2025, SiliconANGLE reports.

Cohere's chief AI officer, Joëlle Pineau, told The Globe and Mail that the first version worked largely as a chatbot that could pull internal documents and handle some automated workflows.

"It was essentially a generative model," she said. "Today, we've actually put in a lot of the features that allow agentic behaviour."

Cohere says a year of large production deployments in finance, healthcare, telecommunications, manufacturing, energy and the public sector shaped the upgrade.

At the center is a new agent harness, a redesigned orchestration system for multistep work.

Pineau's analogy: "The model is like the engine, and the harness is like the vehicle," she told The Globe and Mail. "If you have a motor, you've got a ton of power. But without a vehicle, you're not going anywhere."

The orchestration system works through multistep tasks on its own and pulls in a human for steps that need sign-off, SiliconANGLE reports. Around it, Cohere has added:

• Memory: agents keep context across sessions instead of starting cold every time.
• Skills: reusable capabilities agents can call on, so teams don't rebuild the same logic.
• Libraries: shared knowledge and assets the whole organization can draw from.
• Shareable agents: agents built from a simple prompt can be shared across a company, so other teams don't build the same workflow twice.
• Applications: users can prototype decks, dashboards, documents and lightweight apps from natural language, right in the chat window.
• Automations: prebuilt templates, a drag-and-drop workflow builder and real-time performance monitoring.

Customers can use North 2 to build agents for more complex jobs, such as resolving customer issues, financial analysis and coding, The Globe and Mail reports. Pineau said some of the features have already shipped to customers.

Cohere has expanded North's connectors. The new list includes:

• Slack
• Microsoft SharePoint, OneDrive, Outlook and Exchange
• Jira and Linear
• Notion
• GitHub

Cohere also plans connectors to financial data providers including PitchBook, Crunchbase, Daloopa, FiscalAI, S&P Global and FactSet.

This is the part aimed squarely at finance teams. AI agents can spend a long time working on a task, not always efficiently, which pushes usage costs up, The Globe and Mail notes.

North Admin, the platform's management console, now tracks token use down to the individual user and agent, SiliconANGLE reports. Administrators can:

• Set consumption tiers for users and groups based on request and token rates.
• Set user quotas and caps across the whole company.
• Get alerts that warn teams before a limit is reached.
• Choose models for each team, including models a customer brings from outside Cohere.

Cohere is also counting on its work with Nvidia to cut running costs. It says its models produce more tokens per second per node on Nvidia's Blackwell and Hopper GPUs. Kari Briski, Nvidia's vice president of generative AI, said that helps North 2 "deliver more tokens for less," according to SiliconANGLE.

Cohere is targeting regulated industries and governments, where control over data matters most.

Deployment options: North still runs in the cloud, in a customer's own virtual private cloud, on-premises, in hybrid setups, or fully air-gapped for organizations that allow no outside connection.

Guardrails per agent: administrators can apply guardrails to individual agents. SiliconANGLE reports the screening covers personal information in prompts and responses, as well as attempted prompt injection.

Autonomy policies: agents can be limited to the actions they're authorized to take and required to seek human sign-off for critical decisions.

Certifications: Cohere cites SOC 2 Type 2, ISO 27001 and ISO 42001 certifications.

The timing matters. AI agents have made headlines in recent months for misbehaving and trying to hack into organizations during evaluations, sometimes successfully, The Globe and Mail notes.

Cohere has named Bell Canada, LG CNS and Dell as North customers. It declined to say how many clients it has in total.

"We have dozens of companies, and some of them are up to tens of thousands of users," Pineau told The Globe and Mail.

• Bell Cyber: Bell Canada's security unit uses North inside its own security operations. Jawed Ahmad, Bell Cyber's chief technology and AI officer, said it brings in "the control, privacy, and sovereignty requirements our customers demand built in from the start."
• LG CNS: the South Korean IT services company runs North internally and offers it to customers across Korea.

Cohere is much smaller than its US competitors, The Globe and Mail reports. OpenAI sells ChatGPT Enterprise and is rolling out its new Dots agent, Microsoft sells 365 Copilot, and Mistral is making inroads with Canadian businesses.

Open to other models: because North is model-agnostic, customers can even use it with models from Cohere's rivals. "We actually derive a sizable portion of our revenue from the harness," Pineau said.

Getting bigger: North 2 arrives less than three weeks after Cohere signed a merger agreement with German AI developer Aleph Alpha, SiliconANGLE reports. The combined company will keep the Cohere name and gain more access to the European market.

Training AI models is expensive. OpenAI, Anthropic and Google are spending billions racing to build more capable systems, The Globe and Mail notes.

Cohere, founded in 2019, still trains its own models while building North around them. It focuses on businesses, regulated industries and public-sector clients for whom privacy, security and control are crucial, according to the paper.

"The hypothesis is you do a much better version of both when you develop both," Pineau said.

The Nvidia work is part of that bet. Cohere says its models, trained and deployed on Nvidia's Blackwell and Hopper hardware, help North generate more insights while spending fewer tokens. "Organizations need control over where their AI runs and how their data is secured," Nvidia's Briski said in Cohere's announcement.

A real-world example: Radical Ventures' North automation prepares the briefing its team reads before a founder call. Before it existed, partner Parasvil Patel said, the team spent hours gathering basic context, SiliconANGLE reports.

• For IT leaders: shared agents, skills and libraries could cut down on teams rebuilding the same workflows.
• For finance teams: per-user and per-agent token tracking and hard caps make agent spending easier to predict.
• For regulated industries: air-gapped and on-premises options keep data inside your walls.