Cohere North 2: Enterprise Agents Get a Memory, a Budget Cap, and a Rebuilt Harness

Enterprises deploying agents have long accepted a three-way trade-off: pick two of security, capability, and cost control. On October 5, 2026, Cohere shipped North 2 straight at that compromise. It is a ground-up rebuild of its enterprise agentic platform with three headline changes: cross-session agent memory, a redesigned orchestration harness, and per-team token spending caps, all available across cloud, on-premises, and air-gapped deployments. North has been generally available since August 2025, and Cohere calls this its biggest upgrade yet. Here is what each piece changes, why it matters, and the code underneath, along with the caveats a careful buyer should carry into an evaluation.

The hard part of enterprise agents is governance, cost, and deployment shape, not the demo

Why enterprise agents stall at the deployment wall

According to Cohere, agents prove value quickly in proof-of-concept work but stall in production. The blocker is not model intelligence; it is three missing things: accurate retrieval, persistent memory, and enterprise-grade cost and security governance. Without a single platform, teams stitch together disparate tools that break every time the underlying model or framework changes. Cohere positions North 2 as a rebuild that refuses the compromise: security, capability, and cost control together, not two of three. That is why memory, orchestration, spending caps, and deployment shape all arrive in the same launch rather than four separate point releases.

# North 2's pitch, in one object. Facts are from Cohere's
# October 5, 2026 launch materials and VentureBeat's coverage.

NORTH_2 = {
    "launched": "2026-10-05",
    "adds": [
        "cross-session agent memory",
        "redesigned orchestration harness",
        "granular admin cost controls (token spend caps)",
    ],
    "deploys_to": ["cloud", "on-premises", "air-gapped"],
    "tradeoff_refused": "security vs capability vs cost control",
}

The rebuilt orchestration harness

The core of North 2 is a rewritten orchestration harness. Previously, every team built the plan-route-execute loop itself, duplicating work and drifting apart. North 2 collapses that into one harness: a planner decomposes the goal into steps, a router decides which tool or sub-agent runs each step, an executor calls tools and collects results, a memory module reads and writes cross-session state between steps, and a guard enforces policy and budgets before every tool call. Centralizing this logic means policy, audit, and quotas are issued from one place, teams stop rebuilding core infrastructure, and the question of what an agent is allowed to touch becomes answerable in a single, governed surface.

# The orchestration harness is the part teams had been gluing
# together themselves. North 2 centralizes it.

HARNESS = {
    "planner": "decompose the goal into steps",
    "router": "decide which tool or sub-agent runs each step",
    "executor": "call tools and collect results",
    "memory": "read/write cross-session state between steps",
    "guard": "enforce policy + budgets before each tool call",
}

# Before: every team rebuilt planner/router/executor and drifted.
# After: one harness, governed in one place.

Per-team token budgets and usage, made visible

Cross-session memory without the context tax

The headline capability is cross-session memory. The design goal is explicit: keep the within-session prompt bounded while persisting what actually matters across sessions. Within a session, North 2 keeps an active turn buffer. Across sessions, it stores facts, decisions, and preferences. The crucial detail is retrieval: instead of stuffing full history into the prompt, it runs a ranked lookup and pulls only the matching facts. When a user returns days later, the agent can reconstruct last week's context without paying a token cost that grows linearly with history. For multi-turn, long-horizon work, that is the difference between a flat cost curve and an exploding one. It also changes what teams must build: instead of dumping every prior message into the prompt and hoping the model sorts it out, they define what deserves to be remembered as a fact, and let retrieval decide what re-enters the context each turn.

# Cross-session memory is the headline. The design goal is to keep
# the prompt bounded while persisting what matters across sessions.

MEMORY = {
    "within_session": "active turn buffer",
    "across_sessions": "facts, decisions, preferences",
    "retrieval": "ranked lookup, not full-history stuffing",
}

def build_context(query, memory_store, budget_tokens):
    hits = memory_store.search(query, limit=8)      # relevance, not recency
    facts = [h.summary for h in hits]
    return facts, budget_tokens                     # caller enforces the cap

Cost governance: capping token spend per team

North 2 promotes cost governance to a first-class control. Administrators set token spending caps per team or workspace, on a monthly window, and define what happens on breach: warn, throttle, or block. Metering covers input tokens, output tokens, and tool calls. The practical value is that a runaway agent cannot quietly burn a quarter's budget. Note that Cohere does not publish North pricing; it says cost scales with the size and complexity of a deployment, including infrastructure requirements, usage patterns, support, and customization. Cost control is now an engineering capability, but the bill still needs to be aligned with the vendor per scenario.

# Cost governance is now a first-class control. Caps are set per
# team/workspace, so one runaway agent cannot blow the quarter.

BUDGET_POLICY = {
    "scope": "team",
    "window": "monthly",
    "on_breach": ["warn", "throttle", "block"],
    "metering": ["input_tokens", "output_tokens", "tool_calls"],
}

def before_tool_call(team, policy, spent, estimated):
    if spent + estimated > policy.limit:
        return policy.on_breach[-1]     # block by default
    return "allow"

# Cohere does not publish North pricing; it says cost scales with
# deployment size, usage, support and customization.

Cross-session memory turns last week's context into retrievable facts

Deployment anywhere: cloud, on-prem, air-gapped

For regulated buyers, deployment shape is a governance decision, not a checkbox. North 2 targets three: cloud (vendor-managed, data kept in-region), on-premises (customer-managed, data stays on site), and air-gapped (customer-managed, no egress). The last two are frequently a precondition for purchase in healthcare, defense, and finance, not a nice-to-have. Notably, Cohere announced a partnership with PwC the same day to accelerate secure enterprise AI adoption. Putting deployment shape in the product and bringing in a consultancy signals that Cohere is aiming at organizations that must prove compliance before they can ship, rather than at teams chasing the fastest demo.

# Deployment shape is a governance decision, not an afterthought.
# Air-gapped and on-prem matter for regulated buyers.

DEPLOYMENT_TARGETS = {
    "cloud":      {"control": "vendor-managed", "data": "in-region"},
    "on_prem":    {"control": "customer-managed", "data": "stays on site"},
    "air_gapped": {"control": "customer-managed", "data": "no egress"},
}

# Cohere announced a partnership with PwC the same day to accelerate
# secure enterprise AI adoption.

What it signals for the agent platform market

North 2 makes a broader point clearer: the competition among agent platforms is shifting from whose model is strongest to who can keep agents under control in production. Memory solves continuity, orchestration solves reuse, spending caps solve controllability, and deployment shape solves compliance. Together they are the precondition for putting agents on a critical path. VentureBeat's survey data points the same way: in orchestration control planes, security and permissioning limitations are the top risk at 37 percent, followed by vendor lock-in at 23 percent, and 53 percent expect a hybrid control plane. The next winners are likely vendors that make governance a product core rather than an afterthought. Cohere's bet is that the buyers who stalled at the deployment wall will pay for that governance directly, and that memory, budgets, and deployment shape bundled together are a more compelling purchase than any single capability on its own.