On 19 October 2026, Notion announced the sixth generation of its agent harness — the infrastructure layer that powers Notion Agent and Custom Agents. The headline is time: a single agent task can now run for more than 24 hours, up from five minutes. That is not a marketing number. It is a structural signal about where enterprise agent infrastructure is heading, and what buyers should expect from the platforms they evaluate.
Notion has been steadily upgrading its AI product line since launching its first agent harness. Each version has added model providers, tool integrations and orchestration features. But this release is different because it addresses the constraint that has quietly limited every agent deployment so far: if an agent can only work for five minutes, it cannot complete genuinely complex business tasks. Reconciliation, research, multi-step content pipelines and cross-system data processing all take longer than that. By moving to dedicated compute with queue-backed worker loops, Notion has effectively declared that agent tasks should be measured in hours, not minutes.
What changed under the hood
Notion's co-founder Akshay Kothari summarised the release around three pillars: capability, efficiency and control. Each of those translates into specific infrastructure changes that enterprise buyers should understand.
Dedicated compute with durable execution. Every agent session now runs on its own service with a queue-backed worker loop. If the session restarts, the task resumes from where it left off. That is how the 24-hour runtime is possible. It also means agents survive network interruptions, node failures and updates without losing context. For enterprise IT teams, this matters because resilience is not a feature — it is a requirement for anything that touches production data.
Computer and browser use. Agents now have access to their own computer and browser. Instead of generating scripts with language model tokens, they can write and execute code directly. Notion says this produces faster, cheaper and higher-quality results for tasks that benefit from programmatic execution. For buyers, the implication is that agents are becoming hybrid systems: language models for reasoning and judgement, actual compute for speed and precision.
Smarter context management. Long conversations get compacted using each model provider’s native compaction methods. Large tool results are offloaded instead of flooding the context window. Notion claims this reduces compute per task, and early testing shows Custom Agent runs cost about 30 per cent less. Any cost reduction at the infrastructure level eventually flows through to the buyer’s bill, whether that is per-seat pricing or consumption credits.
A "Review for me" safety mode. A separate model checks whether an action is safe before the agent executes it. This is not full governance, but it is a recognition that autonomous agents need layered safety checks — not just a single prompt asking the model to behave well.
Up to 50 per cent faster on mobile. A performance improvement that sounds like a consumer feature until you consider how many enterprise tasks start on phones: approving requests, reviewing dashboards, flagging exceptions. Speed on mobile is speed to decision.
The Auto router: frontier quality at 90 per cent lower cost
Notion upgraded Auto, its model router that sends each task to the model that best balances quality, speed and cost. Across twenty-plus common knowledge-work tasks, Auto delivered similar results to frontier models at roughly 90 per cent lower average cost. If a model stalls on a task, Auto re-routes to a different provider.
The enterprise implication is clear: model selection is becoming automated infrastructure, not a manual procurement decision. Buyers who invest in platforms with intelligent routing get near-frontier performance without the per-token premium. Over time, this changes the vendor evaluation criteria. It is no longer just which model a platform uses, but whether the platform can route intelligently across multiple models and providers.
Notion added seven new models across Anthropic, OpenAI and open-weight providers in October alone. Auto integrates new releases without requiring buyer action. For IT teams already struggling to keep pace with the model release cycle, this is meaningful operational relief.
Bring your own key, bring your own model
Notion also announced two features that enterprise procurement teams have been asking for. ChatGPT Plus, Pro and Go subscribers can now use their existing plans inside Notion Agent — a direct integration built in partnership with OpenAI. And Bring Your Own Key (BYOK) is in private alpha for Enterprise plan customers, allowing organisations to connect their own model-provider keys for greater control over data residency and spending.
The improved Notion MCP is also significant. MCP is now the fastest-growing product in Notion's history, which means enterprise buyers are actively using Notion as a context layer for external agents including ChatGPT, Claude, Grokbot and Muse. The direction Notion is signalling is clear: it wants to be the neutral infrastructure that any agent can read from and write to, rather than a walled garden with its own limited agents.
Governance infrastructure: the AI Command Centre
The most enterprise-facing announcement is the upcoming AI Command Centre. Admins will get a unified dashboard showing credit usage by member, feature and agent, with configurable alerts and spending limits. They can also choose which models members can access, and control who can create, pause or archive Custom Agents.
This is not revolutionary, but it is necessary. As agent platforms move from individual power users to team-wide and organisation-wide deployment, centralised visibility becomes a hard requirement. Without it, finance teams cannot budget, security teams cannot audit, and compliance teams cannot demonstrate control. Notion is acknowledging what every enterprise buyer already knows: the agent is only half the product. The other half is the infrastructure to manage it at scale.
New credit system: two tiers, clearer billing
Notion is replacing its existing credit model with two monthly allowances: Standard credits for Auto-routed work, and Select credits for work where the user picks the model or runs Custom Agents. Business and Enterprise seats include 1,500 Standard and 500 Select credits per month. The vast majority of users are unlikely to run out, according to Notion.
Admins can purchase additional credits with full allocation control. Custom Agents no longer require a separate credit purchase to start building. The system goes live for new customers immediately; existing customers transition on 2 November 2026.
For procurement, the value is transparency. Separating Auto-routed from user-selected work creates a natural cost baseline and a variable-cost layer for advanced use. It is a structure that maps easily onto enterprise chargeback and budgeting models.
Why this matters for the enterprise buyer
The Notion announcement is best read as a benchmark. It shows what a mature agent platform is starting to look like in late 2026: model-agnostic routing, durable long-running infrastructure, cost transparency, governance controls and open integrations via MCP. These are not premium features. They are table stakes for enterprise adoption.
Buyers evaluating agent platforms should now expect to see:
- Runtime infrastructure that can sustain multi-hour and multi-step tasks without losing state
- Intelligent model routing that optimises cost without sacrificing quality
- Options for using existing model subscriptions or vendor-managed inference
- Enterprise-grade governance dashboards with spend controls and audit trails
- Open integration standards like MCP that prevent lock-in to a single agent ecosystem
If a platform cannot demonstrate these capabilities, the buyer should ask why — because platforms that can are already shipping them.
The Agentic Expo angle
Agentic Expo exists because the gap between agent capability and enterprise readiness is closing fast. Notion is not an enterprise software company in the traditional sense, but its agent harness release sets a reference point that enterprise buyers and enterprise vendors will both be measured against. Long-running tasks, cost-efficient routing, transparent governance and open integration are the dimensions on which agent platforms will compete in 2027.
By March 2027, when Agentic Expo opens at Olympia London, the enterprise buyer will expect these capabilities as standard. Our exhibitors are building them today. The buyers attending will be looking for evidence that the platforms on the floor can match or exceed what they have already seen in the productivity tools their teams are already using.