Picking an AI agent platform used to be about one question: how smart are the agents? Now clients ask a second one, and they ask it loudly. How much will this cost us when hundreds of staff start using it?

On October 5, 2026, Cohere launched North 2, a big update to its enterprise agent platform. The headline features are agent memory, shared skills and something finance teams will like: hard caps on token spending for each user, team and agent.

I'm Nitin Kumar, Lead Project Coordinator at Debugged Pro. We build software, web apps and AI features for clients in India and the USA. I read Cohere's announcement and the coverage from VentureBeat and SiliconANGLE. Here is what changed, what it means for agencies, and how I would approach it with a client.

Cohere's launch graphic for North 2. Image: Cohere

What Cohere Announced With North 2

North is Cohere's workspace for running AI agents inside large organisations. It has been generally available since August 2025. Cohere says a year of large production rollouts in finance, healthcare, telecom, manufacturing, energy and the public sector shaped this new version.

In its official North 2 announcement, Cohere frames the problem simply. Enterprises felt they had to give up one of three things: security, capability or cost control. North 2 is its attempt to offer all three in one place.

The update falls into five groups:

A new agent harness that runs multi-step work on its own and pulls in a human for steps that need sign-off.
More context through memory, shared libraries, reusable skills and app creation.
Model choice, so teams can use Cohere's models or bring their own.
Flexible deployment in the cloud, on premises or fully air-gapped.
Governance through North Admin, with cost controls, rate limits, user quotas and company-wide caps.

The New Agent Harness and Memory

A harness is the layer that tells an agent what to do next, which tools to call and when to stop. Cohere says it redesigned this layer so agents can handle longer, more complex tasks with fewer failures.

Several features sit on top of it:

Skills: reusable pieces of logic that agents call, so staff don't rebuild the same steps again and again.
Libraries: shared documents and assets the whole company can draw on.
Memory: agents keep context across sessions instead of starting cold every time.
Applications: users can create decks, dashboards, documents and light apps from a plain request in chat.
Automations: templates, a drag-and-drop workflow builder and live performance tracking.

Paul Teyssier, VP of Product AI at Cohere, told VentureBeat that memory lets context last longer across sub-sessions and agents. He did not give a number for how much longer. Cohere also did not share performance benchmarks for North 2, saying apps are harder to benchmark than models.

New connectors

North 2 now connects to Slack, SharePoint, OneDrive, Outlook, Exchange, Jira, Linear, Notion and GitHub. Cohere also plans connectors to financial data providers such as PitchBook, Crunchbase, S&P Global and FactSet.

For agencies, this list matters more than it looks. Most client work lives in exactly these tools. An agent that can read a Jira board and a SharePoint folder is far more useful than one that only answers chat questions.

Abstract illustration of connected glass panels in green and purple light, representing agents linked across an AI agent platform — Illustration from Unite.AI's coverage of the North 2 launch. Image: Unite.AI

Token Spending Caps: Why This AI Agent Platform Feature Matters

Agents are expensive in a way chatbots are not. A chatbot answers once. An agent may loop, retry and call tools dozens of times before it finishes a task, and every step burns tokens.

VentureBeat's own research shows the gap. Its VB Pulse survey from July found that 21% of enterprises only monitor agent spend after the fact, with no real-time way to stop an agent before a large bill arrives.

North Admin is Cohere's answer. According to Cohere and VentureBeat, admins can:

See token use down to each user and each agent.
Set consumption tiers for users and groups by request and token rates.
Cap usage across the whole organisation.
Set alerts that warn teams before they hit a limit.
Control which models each team can use.
Break usage down by model, user and department, and export it to their own analytics tools.

Cohere also says its models produce more tokens per second per node on Nvidia Blackwell and Hopper GPUs, which lowers the cost to run. Kari Briski, Nvidia's VP of generative AI, said this helps North 2 "deliver more tokens for less."

How North 2 compares with rivals

Cohere is not alone here. VentureBeat notes that the big players have added their own spend controls:

Platform	Spend control described by VentureBeat
Microsoft Copilot Studio	Monthly consumption limits for each agent
ChatGPT Enterprise	Workspace default, group limits and individual overrides
Claude Enterprise	Limits for the organisation, groups and individual users
Cohere North 2	User and agent tracking, tiers, company caps and alerts, across Cohere and outside models

The difference Cohere is pushing is portability. The rivals' controls govern use inside their own products. North 2 lets you assign outside models and still manage spend in one console.

Security and Private Deployment

Agents touch internal tools, so a mistake can do real damage. North 2 adds guardrails at the level of each agent. They scan prompts and responses for personal data and try to detect prompt injection.

Autonomy policies limit each agent to approved actions and send critical decisions to a person. North holds SOC 2 Type 2, ISO 27001 and ISO 42001 certifications, and it can run self-hosted, in a private cloud, hybrid or fully air-gapped.

One customer quote stood out to me. Jacob Rideout, CTO of HiddenLayer, said: "We only turned those workflows on once we could see and control what the agents were doing." That is the right attitude for any client rollout.

If you want a deeper look at watching agents in production, our post on AI observability and the Dynatrace Arize deal covers the monitoring side.

Who Is Using North Today

Cohere and SiliconANGLE named a few customers:

LG CNS, the South Korean IT services firm, runs North internally and offers it to its own customers across Korea.
Bell Cyber uses North inside its security operations in Canada.
Radical Ventures built an automation that prepares the briefing its team reads before each founder call.

The LG CNS example is the one agencies should study. It is an IT services company using the platform for itself and then reselling that know-how. That is a model many Indian IT firms already follow with other vendors.

The launch also comes less than three weeks after Cohere signed a merger agreement with German AI company Aleph Alpha, according to SiliconANGLE. Cohere is clearly betting on buyers who care about where their data lives.

What This Means for Agencies in India and the USA

VentureBeat's July survey asked which agent platform companies use most. Among clear answers, Microsoft led with 41%, then Anthropic's Claude platform at 28%, LangChain or LangGraph at 10% and OpenAI at 7%. The sample was self-selected, so treat it as a rough signal.

Here is how I read the North 2 launch for agency teams.

1. Cost control is now a sales question

Clients will ask who can cap spend and how fast you can stop a runaway agent. Have a clear answer for every AI agent platform you propose. Show them the dashboard, not just the demo.

2. Model portability reduces lock-in risk

In the same survey, 23% of respondents named vendor lock-in as a top risk, and 37% named security and permissions. A platform that lets you swap models gives you room to move when prices or quality change.

3. Context is where agents fail

VB Pulse found that 68% of enterprises traced a confidently wrong agent answer to missing or messy business context in the past six months. Memory and shared libraries help, but only if the source data is clean. Data cleanup is real, billable work for agencies.

4. Private deployment opens regulated work

Banks, hospitals and government clients often can't send data to a public cloud. On-premises and air-gapped options make agent projects possible in these sectors. If retrieval quality is part of the job, our guide to Cohere Embed 5 for agency teams explains the search side of Cohere's stack.

5. Governance needs a framework

Guardrails in a console are a start, not a full policy. The NIST AI Risk Management Framework gives a practical structure for mapping, measuring and managing AI risk. It works well as a checklist when you write a client's agent policy.

A Simple Checklist Before You Pick an AI Agent Platform

List the workflows. Start with two or three tasks that are repetitive and easy to measure.
Check the data. Confirm the documents and systems the agent needs are accurate and reachable.
Set a budget first. Decide monthly limits per user and per agent before launch, not after the first bill.
Define human sign-off. Mark which actions always need a person, like payments or customer emails.
Test for prompt injection. Feed the agent untrusted content and see what it does.
Compare at least two platforms. Run the same pilot on North 2 and on the client's current Microsoft, OpenAI or Anthropic setup.
Report usage weekly. Share spend, task success and errors with the client during the pilot.

Open Questions to Watch

When will Cohere share benchmark data for North 2?
What will pricing look like for mid-sized companies, since none is published?
Will Cohere add its own model router to send simple tasks to cheaper models?
How will the Aleph Alpha merger change the product for European and Asian buyers?

Key Takeaways

North 2 is a major AI agent platform update, with a new harness, memory, skills and shared libraries.
Token caps are the standout feature, with tracking down to each user and agent.
It is model agnostic, so teams can bring outside models and manage them in one console.
Private and air-gapped deployment keeps it relevant for regulated clients.
There is no public pricing or benchmark yet, so a careful pilot beats a quick switch.

Planning an agent rollout for your business or a client? Our team offers free software development consulting to help you compare platforms, set spend limits and build a pilot that proves value.

About the Author

Nitin Kumar is the Lead Project Coordinator at Debugged Pro. He works with development teams and clients across India and the USA to plan and ship web, mobile and AI projects. He writes about practical AI adoption, agent workflows and delivery for agencies and growing businesses.