Cohere's North 2 puts AI agents on a budget and gives them a memory

Justy So Cohere basically turned North 2 into an AI agent with a corporate credit card and a chaperone.

Cody Yeah, that line about user quotas, rate limits, and org‑wide caps is doing a lot of work. It’s basically, “your agents cannot YOLO the cloud bill anymore.”

Justy And memory. They’re clearly aiming at those VentureBeat numbers: twenty‑one percent of enterprises only reactively watching spend, and sixty‑eight percent tracing bad answers back to missing context. That’s the whole pitch in two stats.

Cody Right. The argument is: if you give agents durable memory across sessions plus hard budgets and approvals, you fix the two biggest enterprise pain points — cost shocks and confidently wrong outputs from thin context.

Justy Which is such an Exploring Next move. We’ve been stuck in this coordination machinery loop since episode nine‑eighty‑one, and Cohere just shipped it as a product.

Cody Mm‑hm.

Justy Although I will say, if you’d told me last November we’d be this excited about “flow control sets consumption tiers,” I would’ve laughed you off a call.

Cody You still kind of are. But yeah, the life we chose. Okay, let’s pick this apart. The redesigned harness is the interesting bit to me: complex multi‑step processes, human‑in‑the‑loop at designated decisions, skills as reusable capabilities, plus shared agents you can publish across the org.

Justy That’s the boring camp again — in a good way. It’s basically, stop copy‑pasting the same tool logic into ten chatbots. Wrap it into a skill, stick it in a library, and let every agent call it.

Cody The libraries and memory combo is the context fix. Shared knowledge and assets everyone can draw on, and then memory to keep user and workflow state across sub‑sessions. Teyssier says context persists longer across sub‑sessions and even between agents, but he won’t give a concrete number, which is slightly annoying.

Justy Yeah, “longer” is doing some heavy lifting there.

Cody Exactly. Technically, I buy the architecture: you centralize state in a store the harness can route to, instead of stuffing everything into the prompt window. But without numbers — error rates before and after, how often memory retrieval actually changes an answer — it’s just design‑doc nice.

Justy The evidence is very survey‑heavy. VB Pulse says only three percent of respondents have zero orchestrated workflows, so almost everyone has agents doing multi‑step stuff already. But a big chunk still can’t control spend in real time, and most have been bitten by bad context. Cohere is basically saying, “North 2 is your control plane for that mess.”

Cody And they lean hard on security and deployment. Run it in the cloud, on‑prem, or fully air‑gapped. SOC 2 Type 2, ISO twenty‑seven‑thousand‑one, forty‑two‑thousand‑one, guardrails scanning prompts and responses for P I I and prompt injection, autonomy policies that gate critical actions to a human… it’s the full checkbox buffet.

Justy Plus North Admin. Roles and permissions, which models are used where, real‑time analytics down to user and agent, token spend, limits, exportable usage by model, user, and department. And flow control as the part that actually stops the agent before it blows the budget.

Cody That’s the piece I like most. The article calls out Copilot Studio, ChatGPT Enterprise, and Claude Enterprise all having spend controls now, but those are fenced inside each vendor’s world. North 2 tries to sit over multiple models — Cohere’s or whatever you already use — with interchangeable model, harness, search, and embedding components.

Justy So the central claim is basically: enterprises don’t just want one vendor’s walled‑garden orchestration. They want a hybrid control plane by, what, end of twenty‑twenty‑six? Over half the survey respondents expect that. Cohere wants North 2 to be that neutral-ish layer.

Cody With the usual asterisk that the survey’s self‑selected and small — sixty‑one clear answers, Microsoft leading, then Anthropic, then LangChain or LangGraph, then OpenAI. It’s enough to say, “this is what early adopters think,” not enough to declare a market share chart.

Justy Totally. From a product lens, the who‑should‑care filter feels pretty tight. If you’re already deep into Copilot Studio or Claude Enterprise, their built‑in limits and workspaces might be “good enough” unless you need on‑prem or air‑gapped deployments, or you’re running multiple model providers and want one harness.

Cody Or you care a LOT about untrusted data. That HiddenLayer quote is doing some narrative work: they’re pulling from marketing, engineering, security ops — basically the worst soup of sources — and they only turned those workflows on once they could see and control what the agents were doing.

Justy Which is exactly the ep nine‑sixty‑eight thing: the safe architecture is where the model can be wrong without the workflow being wrong. Autonomy policies, human approvals, logs of all changes — that’s the stuff procurement and audit will actually ask about.

Cody The missing piece is hard performance data. Cohere tells VentureBeat they’re working on benchmarks, but “applications are harder to benchmark than models,” which is true and also a convenient way to avoid publishing failure rates. I don’t think that kills the argument, but it means North 2 is mostly a governance and deployment story, not an empirical reliability story yet.

Justy Yeah, this doesn’t magically make your agents smart. It just keeps them on a budget, gives them somewhere to remember things, and lets your security team sleep a bit better. If you’re the person who currently discovers runaway spend from a finance email, or you’ve had to roll back an agent because nobody could explain what it did, this is a real change.

Cody And if you’re still in the “one enthusiastic team has a LangGraph thing in a corner” phase, this is probably overkill. You don’t start with a hybrid control plane; you earn it by suffering through enough one‑off agents.

Justy So basically, North 2 is Cohere betting our boring‑camp thesis is the actual market: shared harness, spend controls, memory, and approvals as the main act, models as swappable parts. I’m kind of rooting for that to be true.

Cody Same. Ask me in six months whether anyone outside case studies is actually using the reusable skills and budget caps the way this article imagines.

Justy Deal. Until then, Cody, I’m putting you on a strict token budget for episode ten‑forty‑six. Don’t blow it.