Meta Muse made everyday messages and recurring tasks easier to follow, while OpenAI Dots worked more naturally alongside Slack and Codex. The difference became clearest where a person had to step in: Dots needed help finding the right context for some work and keeping a delegated task in its intended project. Muse offered clearer places to inspect its memory and schedule. Both are persistent cloud agents, meaning they can keep working across conversations and run tasks on a schedule, but their strengths showed up in different workflows.
Sending a message and following the reply
Dots handled a workplace exchange cleanly. After a native Slack connection, it received a request in a team channel, showed that it was processing the request, and replied in a Slack thread. Teammates could also address the agent with a mention in a shared channel. For a group already coordinating in Slack, the useful result was not just an answer: the response stayed attached to the conversation where the request began.
Muse handled a comparable messaging task through WhatsApp. A message sent from a phone received a reply within seconds, and the conversation appeared in Muse’s desktop dashboard. The desktop copy was view-only during that exchange, keeping the phone conversation and dashboard from becoming competing places to reply. Muse also separated other topics into side chats, rather than putting every discussion in one long history.
These were not identical tasks. Slack tested shared team work; WhatsApp tested a personal conversation moving between phone and desktop. Both agents completed these exchanges. The choice depends on where a reply needs to land: in a team thread that colleagues can follow, or in a personal message that remains visible in a central dashboard.
▲ Team threads and synced personal messages
Keeping context and recurring work in order
The agents took different approaches to memory. Muse exposed editable MEMORY.md and SOUL.md files. These are Markdown files, or plain-text documents with simple formatting: one holds lasting facts and preferences, while the other sets the agent’s persona and interaction rules. That gave a user a direct place to review or change what Muse should retain and how it should respond. It did not, by itself, establish that every remembered fact was correct.
Dots automatically drew on existing ChatGPT memory, then built its own context. In one attempt, it initially brought up outdated information. Pointing it toward Codex memory and local files helped direct it to more relevant material. For someone whose work already lives in those tools, that connection can save repeated setup, but the result showed why inherited context needs checking before an agent uses it for ongoing work.
The gap was sharper with project organization. Asked to brainstorm within an existing Codex project, Dots created separate conversation threads in the general Recents area instead of nesting the delegated work under that project. Its explanations of where the task belonged were inconsistent. The practical intervention was to inspect the project and organize the resulting threads manually, rather than assume the agent had filed them where requested.
Scheduling exposed a similar difference. Muse showed scheduled routines and a Heartbeat set to run every 30 minutes, with a run log for checks of connected email and Google Docs. An Approvals area offered a place to review actions requiring sign-off, though it held no pending approvals at the time. Dots could support a daily planning routine set for 8:00 PM Central Time, but managing repeating tasks required finding the scheduled-task controls in Codex or ChatGPT Work rather than relying on the main Dots interface. A user should verify both that a routine is scheduled and where to edit it later.
▲ Scheduled work and project organization
Deliverables and model performance
Both agents could produce deliverables such as documents, dashboards, presentations, PDFs, and code. Muse placed its output in an Artifacts library, where examples included dashboards and PDFs. It also offered podcast creation, including a recurring three-minute news briefing setup. Those interface examples show how work can be found or configured; they do not provide a matched test of the quality of two finished deliverables from the same request.
Dots offered other ways to work, including voice calling through GPT Live and access to a connected computer environment with a browser, terminal, and local files. That access fits a developer workflow, but it does not remove the need to check what happened to a delegated task. A public demonstration of a Dots phone call also failed to produce the expected response, so voice availability alone should not be treated as proof that every call will succeed.
The reported model benchmarks add context, but they are separate from the workflow results above. GPT-6 Astra powers Dots, while Muse Spark 1.3 powers Muse. A token is a small unit of text a model processes or produces.
| Measure | GPT-6 Astra | Muse Spark 1.3 |
|---|---|---|
| Terminal and coding benchmark | 59% | 33% |
| SaaS automation benchmark | 68% | 58% |
| Professional-work rating, GDPval-AA | 1,542 | 1,672 |
| Output generation speed | 51 tokens per second | 184 tokens per second |
Astra led the listed coding and automation measures; Spark 1.3 led the listed professional-work rating and generated text faster. None of those scores settles whether a specific message, scheduled task, or project handoff will be organized correctly.
Which workflow should you test first?
Muse Free allows up to 100 million tokens a week at $0 per month, making it a lower-cost way to try personal messaging and recurring tasks. Dots requires ChatGPT Pro, starting at $100 per month for the first dot. The price difference matters most if the agent is new to your workflow rather than already connected to work you do in Codex, ChatGPT, and Slack.
Start with a task you already understand: send a message through the channel you use, inspect where the reply appears, then schedule a repeat task and find its edit controls. In Muse, review the visible memory files and scheduled-task log. In Dots, confirm that inherited memory is current and that delegated Codex work landed in the intended project. Muse appears easier to supervise for personal routines; Dots offers a better fit for an existing developer workspace, provided someone checks its context and task placement.