The September AI Agent Review: Access, Approvals, and Continuity

An AI agent can have a dazzling launch video and still leave you with the same awkward question: what can I actually hand over to it? September brought two useful answers. Meta introduced a personal agent inside familiar messaging channels. Relevance AI expanded the assistant that builds and operates its automation tools. Meanwhile, an existing assistant's team announced a move to OpenAI. The feature list is only part of the buying decision.

This September 2026 AI agent review covers official announcements from September 1–30, with sources checked on October 2. It also separates products newly found during this review from products launched that month. This is a review of published evidence, not a claim that we tested every workflow or measured a new set of rankings.

The short version

1. Meta Muse: worth watching if you want a personal agent accessible through WhatsApp; check rollout eligibility before planning around it.
2. Relevance AI Invent: relevant for teams building persistent agents and workflows; its September update expands what the builder can create and operate.
3. Continuity matters: a capable assistant can change direction. Check what happens to your work, connections, and data when its team moves on.
4. Newly found is not newly launched: Winston, Aori, and Agentsy illustrate different approaches, but their discovery does not establish a September launch date or completed Atlas admission.

Muse brings the agent into WhatsApp

Meta's September 8 announcement of Muse describes an agent people can message in its own app or directly in WhatsApp. The distinction that matters is action: Meta says Muse can work across apps, use a browser, fill forms, and continue longer tasks after someone closes the app.

That is a more useful promise than another place to ask questions. For a founder, the practical test is whether a request can survive the journey from a message to a finished piece of work without turning into a supervision project.

There are meaningful boundaries in the announcement. Meta says users choose access, can disconnect services, and are asked for approval before sensitive actions such as sending email or making purchases. It describes a dedicated virtual machine and a separate Sentinel agent controlling outbound activity. Those are vendor descriptions of the architecture, not an independent security assessment.

Availability also deserves attention. The announcement describes a US rollout. Someone working elsewhere should verify access rather than treating a global launch headline as a working signup path. Features described as coming later should stay out of today's purchasing decision.

Invent expands the automation builder's job

Relevance AI's September 3 changelog says the new Invent can create and edit agents, tools, workforces, knowledge tables, triggers, and evaluations. It can also run tasks and retrieve analytics, with context across the project.

That combination matters. An assistant that only generates a first workflow leaves you responsible for keeping it working. A builder with access to the surrounding project can potentially help with the less glamorous work: changing the trigger, examining a result, and adjusting the next run. The release documents the expanded controls; it does not prove every generated workflow will be reliable.

There is an administrative prerequisite too. Relevance AI says Invent is enabled per project by an organization administrator. If you cannot find it, that setting is a better first check than assuming the feature is broken.

My preference is to judge this category with one recurring job. Describe the input, the expected output, and the point that needs approval. Then inspect what the product actually created. A confident explanation of a workflow is useful; a workflow you can inspect and maintain is what you are buying.

Check continuity alongside capability

The Atlas's September evidence includes a lifecycle change for Kairos. Its official next-chapter page says the team has joined OpenAI and will pursue its work inside Codex and ChatGPT. The page itself does not give a precise announcement date, so I would not turn the date we checked it into a launch or shutdown date.

This is a reminder to look beyond the impressive demo. Before an assistant becomes part of a business process, work out where the underlying files live, which credentials it holds, and whether another tool can take over the job. Those questions are useful for a small startup's product and a large company's product alike.

I would keep the output of a recurring task somewhere I control. A useful research brief should remain useful if its assistant disappears. A calendar event should exist in the calendar, and a document should remain accessible outside the conversation that created it.

Three candidates worth further verification

These products were missing from the pulled Atlas roster and aliases. They are discovery candidates, not newly admitted recommendations, and we have not established September launch dates for them.

Winston describes a hosted personal assistant in WhatsApp, Telegram, and Signal, with memory, voice-note input, multi-step requests, and connections to tools such as Gmail and Calendar. Its pitch is particularly concrete for households: shared context and separate profiles. That makes it worth checking against the assistant criteria, including what is actually available during signup.

Aori takes a different approach. Its official page describes a local-first desktop and Android application, bring-your-own model credentials, Telegram and WhatsApp connections, persistent memory, and scheduled routines. It is explicitly a public beta. The WhatsApp connection uses an account the user controls, while Telegram uses a bot account. Those are material setup differences, not interchangeable channel checkboxes. Its page already describes a July release, so finding it now does not make it a September launch.

Agentsy focuses on social content. Its official page describes planning, captions, generated visuals, and Instagram publishing, with approve-or-skip replies through WhatsApp. That is a narrower promise than a general personal assistant, and potentially an easier one to evaluate. Further verification needs to distinguish a real agent conversation from an approval-only interface and confirm the persistent workflow and access requirements.

What I would ask before choosing

Start with the work, then choose the interface. If the request begins as a voice note while you are out, messaging access matters. If it starts with an incoming business event, inspect triggers, logs, retries, and approval controls. If it involves private local files, check exactly what stays on the device and what goes to a model provider.

The Notis assistant finder and automation finder separate those needs. Use the underlying evidence and limitations to shortlist tools. This review does not claim a September visibility winner: a fresh, complete comparative measurement was not available for this article.

Frequently asked questions

Did every product here launch in September?

No. Muse's announcement and Invent's update have September dates. Winston, Aori, and Agentsy are products discovered during this review; discovery alone does not establish when they launched.

Does WhatsApp support mean a full assistant?

No. A product may offer a two-way agent, approval replies, notifications, or a connector that needs another system. Read the documented interaction and setup before treating them as equivalent.

Are these new Atlas rankings?

No. This article discusses verified product announcements and discovery candidates. It does not present a completed monthly ranking update or imply that pending candidates passed admission.

September's useful question is simple: can this agent finish your recurring job, show you what happened, and leave the result somewhere you can keep? Answer that before adding another subscription.