Cohere North 2 gives agents memory, shared skills and token caps. The checklist every agent rollout needs
Enterprise agent platforms are converging on the same building blocks. You need them too, whatever tool you build on.
What happened
Cohere launched North 2 on October 5, a major update to its enterprise agent platform. According to Cohere, it adds a redesigned agent harness for multi-step automations, plus:
- Skills: reusable capabilities agents can call, so people stop rebuilding the same logic
- Libraries: shared knowledge and assets the whole organization draws from
- Memory: agents keep context across sessions instead of starting cold
- Automations: templates, a drag-and-drop workflow builder and monitoring
- Connectors: including Slack, SharePoint, Outlook, Jira, Linear, Notion and GitHub
On the control side, North Admin adds roles and permissions, cost controls, rate limits, user quotas and org-wide caps, plus alert thresholds before limits hit. Cohere says it is model agnostic and can run in the cloud, on premises or air-gapped. It also lists autonomy policies that make agents seek human oversight for critical actions.
The Register describes skills as standard operating procedures for the agent, often written by doing a task with the agent and documenting the steps that worked. It adds that access control lists on libraries and skills mean an intern cannot reach proprietary data just by asking.
My take
Most of my clients will never buy North. But its feature list is a useful audit for any agent you run on n8n, Zapier, Make or a CRM's built-in AI:
- Is the agent's procedure written down? A skill is just a documented SOP. If yours lives only in a prompt someone tweaks, it will drift.
- Is shared knowledge in one place? Two agents reading different price lists is how customers get two answers.
- Is memory deliberate? Decide what the agent should remember about a contact, and what it should never keep.
- Is there a spend cap per workflow and per user? Alerts before the cap, not an invoice after it.
- Do critical actions need a human? Refunds, deletions and bulk CRM edits should wait for a click.
If you can answer yes to all five, your agent is ready for real customers.
More posts
- Google's EmbeddingGemma 2 runs search and RAG on device in about 191MB of RAM. Private client data can stay localOct 7, 2026
- Customers' AI agents are getting blocked by human checks and bot defenses. Your client's site may be turning buyers awayOct 7, 2026
- Siena raised $17M to give support, shopping and social agents one shared memory of each customer. Your CRM should do the sameOct 7, 2026
