On October 5, 2026, Cohere unveiled North 2, a rebuilt version of its platform for running AI agents inside large organizations. The update adds a redesigned system for orchestrating multi-step agent work and new administrator controls, including token spending caps. North 2 is designed to work with any model, making it a governance layer rather than a model-specific tool.
What North 2 Brings to Enterprise Agent Operations
Cohere released North 2 as a comprehensive enterprise platform for AI agents that can handle multi-step workflows independently and retain context across sessions. The key additions in this release are:
- Redesigned orchestration engine: A rebuilt system for coordinating multi-step agent workflows, enabling agents to execute complex task sequences without human intervention between steps.
- Token spending caps: Administrator controls that set limits on how much can be spent on tokens, giving organizations budget control over agent operations.
- Lockdown mode with strict ACLs: A security mode that restricts agent access through strict Access Control Lists, limiting what agents can do and access in sensitive environments.
- Model-agnostic support: North 2 works with any LLM provider, so organizations are not locked into a single model when building their agent infrastructure.
- Session-persistent context: Agents retain context across sessions, maintaining continuity in multi-step workflows that span hours or days.
The Token Economics Problem
Token spending is one of the most under-discussed challenges in enterprise AI agent deployment. Each time an agent makes an LLM call, it consumes tokens, and those tokens cost money. When agents run multi-step workflows autonomously, token consumption can scale unpredictably. A single agent executing a ten-step workflow might make dozens of LLM calls, each with varying context lengths and token costs.
Cohere's token spending caps address this directly. Administrators can set budget limits per agent, per project, or per department, ensuring that runaway workflows do not create unexpected cost overruns. This is the kind of operational control that finance and procurement teams need before approving agent deployment at scale.
The cost dimension is particularly relevant as organizations move from pilot deployments with a few agents to production deployments with dozens or hundreds. What seems manageable at pilot scale can become a significant budget line item when agents are running continuously across multiple business workflows.
Lockdown Mode: Security for Sensitive Environments
The lockdown mode with strict ACLs is designed for organizations that need to deploy agents in environments where access control is non-negotiable. In lockdown mode, agents operate with a restricted set of permissions defined by Access Control Lists. This means:
- Agents can only access explicitly permitted tools and data sources
- Actions outside the ACL are blocked and logged
- Administrators can audit exactly what each agent attempted to do
This is particularly relevant for regulated industries, government agencies, and organizations handling sensitive data where the risk of an autonomous agent accessing the wrong system is not just a technical issue but a compliance issue.
The Model-Agnostic Advantage
Perhaps the most strategically significant feature of North 2 is its model-agnostic design. By working with any LLM provider, Cohere positions North 2 as the control plane, not the model layer. This matters for several reasons:
| Concern | Model-Specific Platform | Model-Agnostic Platform (North 2) |
|---|---|---|
| Model switching | Requires rebuilding agent workflows | Switch models without reconfiguring agents |
| Vendor lock-in | Tied to one provider's roadmap and pricing | Free to use the best model for each task |
| Multi-model workflows | Limited or unsupported | Route different steps to different models |
| Cost optimization | Constrained to one provider's pricing | Compare and optimize across providers |
The AI model landscape is evolving rapidly. Organizations that build their agent infrastructure on a model-specific platform risk being locked into a provider whose model may not remain the best choice for every use case. A model-agnostic control plane gives organizations the flexibility to adapt as the frontier moves.
How North 2 Fits the Enterprise Agent Governance Landscape
Cohere's North 2 announcement is part of a broader wave of enterprise agent governance releases in October 2026. Tuskira launched an open-source AI agent runtime gateway on October 1, focused on centralized policy, visibility, and token tracking. Oracle introduced Fusion Claw as a governed runtime for autonomous execution. Dataiku released cross-platform agent management capabilities.
The convergence across these announcements tells a clear story: the enterprise AI agent market is moving from the deployment phase to the governance phase. Organizations that have been piloting agents are now asking how to run them at scale with budget control, security boundaries, and operational visibility. The platforms that answer these questions are the ones that will capture enterprise budgets.
Cohere's positioning is distinct. While Tuskira focuses on the open-source, self-hosted approach and Oracle embeds governance into its enterprise application stack, Cohere offers a model-agnostic control room that works regardless of which LLMs an organization has chosen. For enterprises that want governance without committing to a single model provider, North 2 fills a gap that proprietary platforms leave open.
