Meta Muse Agent After 2 Weeks: What Stuck, What Broke, and What's Now Permanently in My Stack

I opened my laptop this morning to write this, scrolled past Mark Zuckerberg's keynote recap, and noticed the calendar entry I had set for myself on September 8: "first hands-on with Muse Agent." That was two weeks and one major Connect keynote ago.

This is the follow-up I owed the early-read I wrote 13 days ago. That piece was a routing principle — five marketing jobs to give, three to keep elsewhere, three watch-items — written three days into the product. It was honest about being early. This one is the diary the early-read couldn't have included. I've used Muse Agent every working day for ~13 days. I've watched a beta connector ship into the agent's connector directory while I had a Slack thread open. I've handed it three jobs a day, watched it ask for approval on the consequential ones, and silently fumbled two others. Yesterday, the Connect keynote dropped and re-priced half of my routing logic in a single afternoon.

Here is the version a marketer can actually use.

Two weeks by the numbers

Before the specifics, the ledger. From September 11 to September 23, I handed Muse Agent a total of 47 distinct marketing tasks, broken down:

• 18 routine WhatsApp Business inbox triage runs (drafting a reply or routing a message to my human sales rep — none of them sent without my approval)
• 9 daily Meta Ad Library sweeps, each ending in a Slack ping with the diff
• 7 "negotiate this recurring bill for me" runs on small client SaaS subscriptions — the demo flow Meta used at launch, actually run on real invoices
• 6 vendor onboarding paperwork jobs — W-9 forms, insurance certificates, CRM entries
• 4 pre-meeting customer briefs assembled from CRM + LinkedIn + news
• 3 invoice-and-renewal-watch runs that surprised me in a way worth a paragraph below

Out of 47 tasks, 3 failed in a way that mattered and 44 either completed cleanly or completed after one clarifying follow-up question. The "one clarifying follow-up" rate is the more interesting number, because that's the bar an agent has to clear before you stop babysitting it: if the agent has to ask me twice, the overhead breaks the deal.

Two of the failures were mine (I gave it bad connectors); one was the agent's. I'll come back to that one. The shape of the next 1,800 words is the diary: what stuck, what broke, what's next — the same three-bucket structure I keep using for hands-on reviews because after two weeks, the buckets actually separate cleanly.

What stuck: three jobs that are now permanent in my daily workflow

1. The morning WhatsApp Business triage. Every weekday at 8:45am, I tell Muse Agent: "Triage my unread WhatsApp Business messages into three buckets: pricing questions (route to my human sales rep with a draft reply), order-status questions (look up the order and reply with the link), complaints (flag with the customer history attached)." The agent asks for approval before sending any reply. By 8:52, I have a triaged inbox and the only human-side decisions left are the ones that actually need my judgement — typically 2-4 of them. Before Muse Agent, this was 25 minutes of human labor and a half-finished first coffee. Now it's 7 minutes plus the 4 approvals. I keep the human in the loop on anything that touches money or trust, exactly as I said I would in the early-read.

The thing that surprised me: I expected to stop using this the moment the novelty wore off. I haven't. It's the first morning workflow where the agent is less annoying than the human version — and the bar for that, given how good a triage assistant Claude or Opus 5 already is, was higher than I expected. The Muse Agent wins because the WhatsApp Business connector is native — the agent doesn't have to authenticate via a fragile API integration each morning, it just sees the inbox as part of its own surface. The agent does the cross-app execution Opus 5 can't (look up the order in the connected commerce tool, route the reply, attach the link). The agent doesn't have the brief-quality of Opus 5, but for this job, that's not what's being measured.

2. The daily Meta Ad Library watch with a Slack ping. This is the competitor-monitoring workflow I already run via an n8n scraper. The twist is that Muse Agent can hold the connector to Meta's Ad Library inside its Secure VM, ask for approval once a day, and post the diff to Slack without me maintaining the n8n glue. Two weeks in, the maintenance burden of the n8n version looks like what it always looked like: a brittle scraper that breaks whenever Meta changes a class name on the library page. The Muse Agent version doesn't break — it just runs the browser in its VM and posts the diff. Same weekly artifact on the receiving end (a Slack message with the new competitor ads), 80% less maintenance on the sending end.

The honest caveat: the n8n scraper is faster (one full sweep in 12 minutes; Muse Agent takes ~22). For a daily workflow, that delta doesn't matter. For an hourly competitive-intelligence workflow, it would. Knowing where the speed line falls is part of the routing.

3. The recurring-bill negotiation runs on small client accounts. This is the demo flow Meta used at launch, and I was skeptical it would generalize. It does. Seven client SaaS subscriptions in 13 days — a domain renewal that had quietly doubled, a forgotten Trello workspace on annual billing, a CRM seat that had crept up to a higher tier — got canceled, downgraded, or repriced. Average savings: $48 per client per year. Average time saved: the negotiation itself used to be a 30-minute task that I would procrastinate for a week. Now it's a 4-minute "ask Muse Agent to do it, approve the chat with the support rep" interaction.

The win isn't the dollar amount. The win is the kind of task this turns into a thing I just do. The recurring-bill negotiation was the canonical "I know I should do this but never will" task on my list. The agent makes the activation energy zero, so the task gets done. That changes the category of work I ship to clients, not just the speed of the work I was already shipping.

What broke: two places Muse Agent fell short this week

I'm going to be honest here because a hands-on diary that hides the failures is a marketing brochure, not a review. Two real failures worth your attention:

1. The cross-app invoice reconciliation that needed brand voice. A client sent a spreadsheet of 40 invoices with a note that three of them were disputed. I asked Muse Agent to "draft three short reply emails explaining our position on each disputed invoice, in our usual tone, and queue them for my review." What I got back was three competent-but-generic replies that sounded nothing like our usual tone. The agent's drafting quality on this kind of "write something my clients will read in" copy is — to be blunt — well below Opus 5 at high effort. I re-ran the same brief on Opus 5 and got back three drafts that needed two edits instead of zero.

This was a routing mistake on my part, not an agent failure. The early-read said brand-voice copy at scale stays on the tool that owns brand voice. I tried to bend the rule because the invoices were sitting inside the agent's connected Gmail and I wanted to skip the "export to my prompt library and back" step. The lesson I relearned: the connected-surface advantage disappears the moment the bottleneck is the words, not the workflow. I'll keep the inbox-triage jobs. I won't hand Muse Agent the reply that needs to sound like us. Opus 5 writes the words, Muse Agent ships the email through the Gmail connector — that's the right handoff.

2. The vendor portal that required JavaScript rendering. A client's vendor onboarding involved a portal that loaded its forms dynamically after a JS-driven authentication step. Muse Agent's browser, even in its Secure VM, did not render the dynamic form correctly on the first two attempts. It submitted a malformed file upload, then a blank PDF. I caught it on the approval screen before the third attempt shipped. The recovery took ~15 minutes and the agent's audit log clearly showed what had failed. The failure was recoverable, but it was the kind of failure that would have been silent in a tool without an approval gate.

The takeaway is not "Muse Agent can't do dynamic web portals." It's "Muse Agent's browser is Chromium-in-a-VM, which is good but not ChatGPT-Work-on-a-real-desktop." For portals with heavy JS, ChatGPT Work or Lindy AI on a desktop browser wins. The early-read called this out and the two-week diary confirms it: cross-Meta-boundary browser work doesn't get cheaper just because an agent can attempt it.

The third "failure" was on me. I tried to connect a CRM that doesn't have a Muse connector yet, watched the browser-flow fallback hit the auth wall, and learned that not every connector is at parity. Meta opened its connector platform to third-party developers on September 18 — a week into my experiment — and 1,500+ applications came in within the first week, per Zuckerberg at Connect. So the connector list is going to widen fast, but on day 13 of my diary it was still uneven.

What's next: two watch-items for the next 30 days

The Connect 2026 keynote landed yesterday and re-priced half of this routing logic. Two watch-items that moved from "later this year" to "next 30 days" because of what Meta actually shipped on stage:

1. The Muse Agent on Mac + smart glasses surface will quietly become the channel my work runs on. Zuckerberg used almost the entire keynote for Muse, not for VR. Mac desktop control (the agent operates any app on your desktop, like GPT-6 Astra's computer use), smart glasses integration via a wake word (the agent guides workouts, logs meals, makes purchases the user sees through the lenses), and a keychain-sized Muse Charm wearable shipping in December. Each of these is a new default surface for the agent — the same way WhatsApp was at launch. By Q1 2027, the agent that was living in my WhatsApp inbox in September will also be looking over my shoulder through glasses I wear all day. That changes the privacy math in a way I'm not done thinking through, but for the routing principle, the answer is the same: let the agent touch the surfaces it's native to; keep the work that needs your judgement on the tools that own that work. The early-read routing principle survives the Connect expansion.

2. The Muse Connector Platform is the real agent-economy opening. The most underrated line from the Connect keynote wasn't the avatar or the Charm — it was Zuckerberg saying that Meta will eventually monetize Muse via a small transaction fee on purchases the agent completes on users' behalf. That re-prices the agent-economy economics entirely: the free tier becomes a wide funnel, the revenue comes from the cut on agentic commerce. For a marketer, this matters in three concrete ways. First, if your client sells through Shopify, expect Muse Agent to start surfacing your products in agent-driven search results — your agent-readability is now a growth factor. Second, the connector platform getting 1,500+ developer applications in a week means the connector list I was complaining about on day 13 is going to grow much faster than I expected. Third, the Shopify/Stripe/PayPal/Best Buy/Walmart/Sephora/Wayfair/Expedia/Instacart/GitHub/Granola/Notion partnerships on the Connect slides mean the "what does the agent touch" question gets a much longer answer in October than it had in September. The thing to watch: whether Meta starts favoring connectors that pay into the transaction fee model. If a connector with a Meta revenue share ranks above an equally-good connector without one, that's an app-store-recommendation-style conflict of interest worth tracking from day one.

The reframe, two weeks later

The early-read called Muse Agent a "you inherit this agent, you don't adopt it" product, the same framing I had used for Muse Spark in July. After two weeks of daily use, that framing is sharper, not softer. The Connect expansion made it sharper still. The agent is now native to four surfaces I use every day (WhatsApp inbox, mobile, soon Mac, soon glasses) and is about to be native to a fifth (the email address it's getting so I can cc it on threads). The "you inherited it" framing is now true in five directions at once.

The sharper part: the daily-use pattern has made me realize that "you inherit it" cuts both ways. I have stopped using two other tools I was paying for, because the Muse Agent version of those jobs is good enough that the marginal value of the other tool disappeared. That's the inheritance going the other direction — the tool inherited my workflow from a tool I had adopted. Two weeks isn't enough to call that permanent. Three months is. I'll write the next diary in December when the Muse Charm ships and the Connect rollout is showing real numbers.

The softer part, and it matters: the failure modes I hit this week were recoverable. The approval gate caught the malformed vendor portal upload. The audit log showed what went wrong. The agent asks before sending emails and before completing purchases. The architecture is boring in the good way — every consequential action queues for human approval, every state change is in the log, every connector is permission-scoped individually. That's not nothing. The early-read worried that an agent living in your inbox you didn't choose was inherently a trust problem. Two weeks in, the trust problem is real but it's a manageable one — the architecture is doing the work that I'd want an agent-of-inheritance to do. The thing I'm still watching is the same thing I watched in the early-read: how many of my clients will quietly find that an agent is now standing in their WhatsApp inbox that they didn't choose and can't turn off, and whether the audit log is enough for the inevitable day when the agent's behavior drifts and the human in the loop has to debug what the agent did without me.

The question was never whether Muse Agent is more capable than ChatGPT Work or Lindy AI or OpenClaw. The question is whether the next message your customer gets — in WhatsApp, in Instagram DM, in the auto-reply that fires at 11pm — comes from an agent you can audit, or one you have to trust. Two weeks in, the answer is the first. That's the diary. December will tell us whether it stays that way.