The launch in brief
Cohere’s North 2 is a platform for running agents, not a new model. It shows enterprise AI competition moving from model quality to the controls around the model: memory, permissions, spending limits and where the system runs. The Toronto company announced it on 5 October 2026, less than three weeks after agreeing to merge with Germany’s Aleph Alpha.
“The model is like the engine, and the harness is like the vehicle,” Cohere chief AI officer Joëlle Pineau told The Globe and Mail. She said the company already earns “a sizable portion” of its revenue from the harness layer.
Companies that commission custom AI agents for internal workflows can now ask any vendor, Cohere or not, for a specific set of features. Can agents keep context safely between sessions? Can spend be capped per user and per agent? Can the system run inside our network?
What is new in Cohere North 2?
The main change is a rebuilt orchestration harness. This is the layer that breaks a request into steps, calls tools and connectors, and decides when a human must approve before the agent continues. Cohere says it is designed for longer multistep work such as resolving customer issues, financial analysis and coding.
Agents now have memory that persists between sessions, so users do not have to re-explain context each time. Teams can also package repeatable logic as skills and share documents and assets through common libraries. A drag-and-drop workflow builder and prebuilt templates target recurring processes. On the security side, Cohere lists PII screening, prompt-injection detection and autonomy policies that limit what an agent may do on its own.
North 2 is model-agnostic. Customers can run Cohere’s own models or connect models from other developers. That is a deliberate contrast with coding agents tied to a single lab’s models.
How do the token caps work?
The new North Admin console tracks token use per user and per agent. Administrators can set consumption tiers by request and token rate, apply organisation-wide usage caps and get alerts before a limit is reached. VentureBeat described the idea as putting “AI agents on a budget”.
This responds to a real problem. A user’s chat session costs roughly the same every time. An agent loop that retries, re-reads documents or fans out to sub-agents can burn through tokens with no fixed upper bound. Without per-agent metering, finance teams see the bill only after the month closes.
Why does Cohere bet on the harness?
Cohere competes with OpenAI, Anthropic, Microsoft 365 Copilot and Mistral for the same corporate budgets. It also has fewer resources to train frontier models. Pineau argued that developing the model and the harness together improves both. Cohere is still putting its commercial weight on the layer that regulated buyers care about most: deployment inside their own perimeter, audit and cost control.
The planned merger with Aleph Alpha points the same way. It gives Cohere a stronger European presence, where sovereign and air-gapped deployment is often a requirement in public-sector and financial-services tenders, not a nice-to-have.
What it means for US & EU software teams
Budget controls should be designed in from the start. Whatever platform you choose, per-agent metering, hard caps and pre-limit alerts are cheap to add at design time and painful to add after the first surprise invoice. They belong in the architecture, next to logging and access control.
Memory is a data-protection question. Context that persists across sessions is personal data under GDPR as soon as it holds anything about an identifiable person. Retention periods, deletion on request and per-tenant separation need to be specified. For EU teams this ties into the record-keeping duties of the EU AI Act for higher-risk uses.
Model-agnostic harnesses reduce lock-in. If the orchestration layer can switch models, you can route cheap tasks to small models and keep frontier models for hard steps. This only works if prompts, tools and evaluations are kept separate from any one vendor’s API.
Air-gapped options change build-versus-buy. Regulated FinTech and HealthTech teams that ruled out agents because data could not leave their network now have off-the-shelf options. That gives them a benchmark for any custom build.
What to do now
- Measure agent spend at the right level. Track tokens per agent and per workflow, not just per API key, before agents reach production.
- Set hard caps and alerts per user, per agent and per organisation, and test what happens when a cap is hit mid-task.
- Write a memory policy: what agents may remember, for how long, how a user deletes it, and how it is separated between tenants.
- Keep the harness model-agnostic by abstracting model calls and maintaining your own evaluation set, so you can switch providers on cost or quality.
- Choose the deployment perimeter first. Decide whether you need SaaS, your own VPC or on-premises before you compare agent platforms.
Frequently asked questions
What is Cohere North 2?
North 2 is the second version of Cohere’s enterprise agent platform, released on 5 October 2026. It is not a new model. It is the layer that runs agents: orchestration, memory, connectors to business tools, security policies and admin controls.
Can North 2 use models other than Cohere's?
Yes. Cohere describes North 2 as model-agnostic. Customers can use Cohere’s own models or plug in models from other developers within the same orchestration harness.
How does North 2 control AI agent costs?
The North Admin console tracks token use per user and per agent, supports consumption tiers by request and token rate, applies organisation-wide usage caps and sends alerts before a limit is reached.
Can North 2 run without internet access?
Yes. Cohere says North 2 can run in the cloud, in a customer’s virtual private cloud, in hybrid setups, on premises or in a fully air-gapped environment with no external network access.
Sources
Cohere — North 2 product page (October 2026)
The Globe and Mail — Cohere unveils latest AI platform with more agentic features for businesses (5 October 2026)
VentureBeat — Cohere’s North 2 puts AI agents on a budget and gives them a memory (5 October 2026)
SiliconANGLE — Cohere unveils North 2 AI agent platform with rebuilt orchestration and token spending caps (5 October 2026)