North 2: What Cohere Announces and What Still Needs to Be Verified

An agent platform with more control, according to Cohere

Cohere presents North 2 as an update to its enterprise AI platform. The offering brings agents, workflows, and security and cost controls together in one environment. Rather than only responding to a prompt, an agent can take part in a sequence of tasks that involves coordinating actions and tools. That is the shift Cohere is proposing: bringing agents closer to ongoing business processes while giving organizations more ability to govern them.

The important distinction is between the product’s intended direction and capabilities that have been verified. Describing a platform as a control center does not, by itself, explain which permissions can be configured, how spending limits are enforced, or what activity is logged. The available documentation confirms that Cohere advertises cost controls, security features, and an agent platform. It does not provide an independent assessment showing that those controls are effective in every enterprise scenario.

It is also worth distinguishing an announced feature from its availability and scope. The information reviewed does not establish which components are enabled for all customers, whether some capabilities depend on a particular configuration, or what technical or commercial requirements apply. North 2 should therefore be understood here as a product announced with certain features, not as a solution whose operational effectiveness has already been demonstrated.

Multi-step workflows: potential value, evaluation still pending

North 2’s description includes agent-driven management of multi-step workflows. In practice, such a workflow might chain together tasks such as retrieving information from an authorized source, summarizing it, and preparing an output for review. This is an example of a possible use, not a demonstration that the product reliably executes every stage or can complete actions without supervision. The quality of the result would depend, among other things, on how the workflow is designed, which tools are connected, and what permissions are granted.

To assess real-world behavior, it is necessary to observe the entire process, not just the final response. A useful test would record which steps were executed, what information the agent consulted, which actions it attempted, and when it requested human intervention. It would also check how the system responds to incomplete data, conflicting instructions, or an unavailable tool. Without such results, the product description cannot establish the level of autonomy or error rate.

Organizations should start with tasks that are limited in scope and reversible. If an agent prepares a proposal, a person can review it before it is sent; if it changes data or initiates an operation, the test should include limits and a recovery path. This approach makes it possible to measure whether automation saves work without confusing the ability to chain steps with authorization to carry them out.

Memory and continuity between sessions

Continuity of context between sessions is among the capabilities attributed to North 2 in the available information. Without sufficient operational documentation, the word “memory” can refer to different behaviors: retaining preferences, retrieving details from an earlier conversation, or preserving a workflow’s state. These are not equivalent, and each raises different requirements for privacy, control, and reliability.

Assessing this capability requires specific answers: what information is stored, for how long, where it is stored, who can access it, and whether it is used for other tasks. It also matters whether an administrator can inspect, correct, and delete context, and what happens after deletion is requested. The evidence provided does not resolve these details, so it is not possible to confirm how memory is implemented or what control options customers have.

A practical test should use synthetic data or information that is not sensitive. Start a task, end the session, and then check what context the system retrieves. Next, try to delete that information and verify that it does not reappear in later sessions or in logs where it does not belong. Repeat the test with users who have different permissions, because memory that helps maintain continuity can also become a route for exposure if account separation fails.

Costs, permissions, and external models

Cohere includes cost governance among North 2’s components. That description is not enough to determine whether budgets are hard limits or alerts, or at what level they can be set: per agent, person, team, or workflow. Nor does it establish what happens when a threshold is reached—whether a task stops automatically or can continue to incur costs. Those differences are essential to understanding whether a control prevents spending or merely reports it.

An evaluation should set small budgets in a test environment, bring them close to the threshold, and then reach it while observing what happens. It should also check whether costs are attributed to a task or user, whether administrators can change the budget, and whether those changes are logged. Written policy and actual behavior should match; otherwise, the limit should not be treated as an effective safeguard.

The available coverage also raises the possibility of connecting models from other providers, but the information that can be verified here does not identify a specific list of models or the conditions for compatibility. Saying that a platform supports different models does not clarify whether every feature works with each one, what data is sent to the provider, or who is responsible for data retention or use. Businesses should confirm these points for each integration rather than assume compatibility is universal.

Security calls for the same level of precision. To decide whether an agent is suitable for production, an organization needs to know how permissions are assigned, which tools it can invoke, which actions require approval, and what logs are retained. The available documentation does not confirm those details for North 2. It is therefore not possible to claim that its controls eliminate the risk of unauthorized actions.

What is still needed to assess North 2 in production

The available information makes it possible to describe the product’s intent: bringing agents, workflows, and enterprise controls together in one platform. It does not make it possible to measure reliability, verify continuity of context, or determine whether spending limits and permissions are enforced. Nor does it provide comparative results or independent tests that would make it possible to assess performance against alternatives under equivalent conditions.

Before a broad deployment, an organization should request technical documentation that applies to its configuration and run a pilot with defined success criteria. These could include the proportion of tasks completed correctly, the frequency of human intervention, permission errors, spending per workflow, and compliance with retention policies. Results should be reviewed across different users, task types, and failure cases; one successful example is not sufficient evidence.

The cautious conclusion is limited: North 2 is presented as an enterprise agent platform with cost and security controls, and the available information describes multi-step workflows and continuity of context as part of its proposition. The extent of real-world control, how memory is implemented, and which external models are specifically compatible remain unestablished. These points should be confirmed through applicable documentation and reproducible tests before the announcement is treated as a guarantee of production behavior.

Keep exploring

Corrections and transparency

If you spot incorrect or outdated information, send us a correction with the page and source we should review.