Notion Auto Launches: Agents Can Run 24 Hours, Costs Slashed by 90%

Notion recently launched a new product called Notion Auto on Product Hunt, rebuilding its Agents from the ground up. The core selling points are straightforward: tasks can run continuously for over 24 hours, automatically route to the optimal model, and cost only about 10% of Opus 5.5 and Astra. For users long plagued by "Agents that quit halfway through" and "token bills that are hard to stomach," these two numbers are compelling enough.

Agent Harness Rebuild: From Sprint to Marathon

The technical core of this update is the rebuild of the Agent Harness. The harness is the execution framework wrapped around the model, responsible for dispatching tools, maintaining state, handling errors, and deciding when to stop. In the past, most Agent products were limited by context windows and session lifecycles, so tasks could be interrupted after a few dozen minutes or require manual resumption. After Notion rewrote the harness, the lifecycle of a single task was extended to 24 hours or even longer.

Supporting capabilities include an operable computer and a browser, which means the Agent is no longer just "reading and writing Notion pages" but can open web pages, click, fill out forms, and scrape information in real environments. This lets it handle tasks closer to real workflows, such as collecting materials across websites and syncing data between multiple systems.

Agent Harness The term originates from the software testing field's "test harness," referring to the scaffolding layer that wraps the code under test and is responsible for starting, driving, and collecting results. In the LLM Agent context, the harness is the execution middleware between the model and the real world, typically including modules such as the task planning loop, tool/function call dispatch, state and memory persistence, retry and error recovery, termination condition determination, and logging and observability. The model itself is only responsible for single-step reasoning; the harness determines how the output of that step becomes an action and how the action result is fed back into the context. Therefore, the same model paired with different harnesses can differ severalfold in actual task completion rate and continuous runtime duration. Current mainstream implementations include the ReAct loop, the two-stage Plan-and-Execute architecture, and long-task schedulers with checkpoints and resumable execution. Notion's emphasis on "rebuilding the harness" is essentially engineering optimization in state management and context compression strategies, rather than a leap in model capability itself.

Review for Me: Adding a Safety Valve for Long-Running Execution

The longer an Agent runs, the greater the risk of errors and abuse. To address this, Notion introduced the Review for me safety mode, which inserts a manual confirmation step before critical operations. This design approach aligns with the current industry discussion on Agent safety: fully autonomous long-running execution still needs supervision points in commercial scenarios, otherwise a single erroneous operation could trigger cascading consequences.

The specific granularity of Review for me (which operations trigger confirmation, whether rules can be customized) was not detailed in the original text, and this is a part that needs key verification in actual use.

Model Routing: Not Bound to a Single Model, Picking the Optimal Solution per Task

Another change in Auto is automatic model routing based on task. Users no longer need to manually choose which model to use; the system dispatches according to task type. Notion claims this scheduling can match the level of Opus 5.5 and Astra while reducing costs by about 90%.

This claim needs to be broken down. The 90% cost reduction is more likely to come from a combined strategy of "using cheap models for simple tasks" rather than a single model being ten times cheaper. The practical significance for users is that high-frequency, low-complexity tasks no longer burn through large amounts of quota, and the budget can be reserved for parts that truly require strong reasoning. As for the conclusion of "matching Opus 5.5 and Astra," it is a vendor's own statement and lacks third-party benchmark verification.

Model routing is a scheduling strategy that routes inference requests before they reach the model: a lightweight classifier, rules, or heuristic scoring determine task difficulty, required context length, whether tool calls are involved, and other features, then dispatch the request to models of different sizes or from different providers. Common approaches include confidence-based cascading (try a small model first, escalate to a large model on failure or low confidence), allocation by token budget, and hard-coded mapping by task type. The benefit of routing comes from the nonlinearity of the cost curve—simple tasks can achieve a close pass rate with cheap models, and only a few difficult problems require strong reasoning models, so the overall weighted cost drops significantly. The cost is that routing misjudgments bring quality fluctuations, and the coexistence of multiple models increases latency and observability complexity. Vendors' claims of "comparable performance at 90% lower cost" usually hold as weighted average results under exactly this kind of combined strategy.

Ecosystem and Billing: Connecting ChatGPT Plans, MCP Overhaul

Notion allows users to connect their own ChatGPT plans, which is a benefit for users who already have subscriptions, avoiding duplicate payment. At the same time, Notion MCP has been substantially rewritten. MCP (Model Context Protocol) is the interface standard for communication between external tools and models; after the rewrite, the stability and coverage of the Agent's calls to external services are expected to improve.

The billing method has also been unified: the new credit system covers all AI usage. Unified credits simplify billing, but they also mean users need to re-estimate their usage costs, especially since long-running tasks will continuously consume credits.

MCP (Model Context Protocol) is a standardized protocol open-sourced by Anthropic at the end of 2024, used to let model hosts discover and invoke external tools, data sources, and prompt templates in a unified way. It adopts a client-server architecture: the MCP Server exposes three types of primitives—resources (readable data), tools (executable functions), and prompts—while the MCP Client runs on the Agent host side and is responsible for connection, authentication, and call forwarding. Compared with the past where each application implemented its own plugin interface, MCP allows the same tool service to be reused by different Agent products, reducing integration costs. The protocol is usually based on JSON-RPC 2.0 at the bottom and supports both local stdio and remote SSE/HTTP transports. Notion's rewrite of its MCP implementation means that its Agent's handshake, error handling, and permission boundaries when calling third-party services will be more standardized, and it will be easier to connect to the existing community MCP ecosystem.

Who Should Pay Attention

Notion Auto is currently only open to new customers, and the original text did not mention when existing users will be able to use it. From its positioning, it targets teams that need Agents to handle long-cycle, multi-step tasks—scenarios such as research organization, competitor monitoring, and content pipelines. If you previously gave up on Notion Agent because of cost or task interruptions, this is worth re-evaluating; but "24-hour runtime" and "90% cost reduction" are both official claims, and the actual results still need to be tested with real workloads.

At a time when Agent products generally compete on single-task capability, Notion has chosen to pull the competition dimension toward continuous runtime duration and cost per task, a direction closer to the real considerations of enterprise procurement.

One-line Summary

Notion rebuilds the Agent Harness, enabling tasks to run for 24 hours with costs reduced by about 90%.

Paragraph Summary

Notion launched Notion Auto on Product Hunt, rebuilding its Agents from the ground up. The technical core is the rebuild of the Agent Harness, extending the lifecycle of a single task from tens of minutes to over 24 hours, and equipping it with an operable computer and browser so the Agent can click, fill out forms, and sync data across systems in real web environments. On the safety side, it introduces the Review for me mode, inserting manual confirmation before critical operations. Model scheduling has changed to automatic routing by task, and the company claims the performance matches Opus 5.5 and Astra while costing only about 10% of them. On the ecosystem side, it supports connecting users' own ChatGPT plans, and Notion MCP has been substantially rewritten to improve the stability of external tool calls, with billing unified into credits. The product is currently only open to new customers, and both the long runtime and cost reduction are vendor claims lacking third-party verification.