One Year of Always-On Agents: What to Watch After Dots Ships

By the close of September 2026, three of the industry's biggest labs had each shipped an agent with its own persistent cloud computer: OpenAI's dots, launched September 29; xAI's Grok Bot, in beta since August 11 and extended to businesses in early September; and Meta's Muse, launched September 8 and expanded to small businesses on the same day dots arrived. Three companies, three products, one three-week window — a genuine collision, not a coincidence of PR calendars.

A year from now, most of what made that collision exciting will be forgotten. Nobody will remember the dot.com domain redirect as anything but a footnote, and the specific feature checklists compared breathlessly in launch-week coverage will have been rewritten by a dozen incremental updates. What will actually determine which of these products — or which combination of them — survives as infrastructure rather than novelty comes down to a handful of questions none of the three companies has fully answered yet.

The pricing question will force honesty

All three companies launched without a complete public price list for anything beyond the free or bundled tier — dots' additional-dot pricing is undisclosed, Grok Bot's business pricing sits behind a waitlist, and Muse's paid-tier thresholds aren't detailed. That's a reasonable posture for week one. It's not a sustainable posture for month twelve. Within a year, each company will have to publish real numbers, and those numbers will reveal something none of the launch keynotes did: how expensive it actually is to run an agent continuously, at scale, across millions of users — a compute-metered cost structure that behaves nothing like a flat SaaS seat price. Whichever company prices this most honestly, rather than most attractively, will earn the trust of the enterprise buyers who matter most for durable revenue.

The safety documentation gap will close, one way or another

Grok Bot's approach — signing into your tools with your own credentials, no confirmed formal safety framework in the sources available at launch — is the highest-risk of the three permission models covered earlier in this series. That gap will not stay open for a year. Either xAI publishes a real safety and data-handling framework as Grok Bot moves from beta toward enterprise general availability, or an incident forces the conversation before xAI is ready to have it on its own terms. Meanwhile, OpenAI's Custom Rules and Auto-Review, and Meta's Sentinel-gated Secure VM, will get their first real stress tests once millions of users — not launch-week early adopters — start running these agents against messy, real-world tool access. The permission models that looked elegant in a launch announcement will either hold up under actual misuse and edge cases, or they won't, and a year is enough time to find out which.

The oversight-fatigue problem will surface everywhere

Every one of these products ships some version of a human-review layer — OpenAI's Activity View, Meta's audit trails, whatever Grok Bot's undocumented "approval for certain tasks" turns out to mean in practice. All three face the same underlying problem: an oversight tool only works if a human keeps using it, and the entire commercial promise of an always-on agent is that a human's attention shouldn't have to scale with what it does. Expect the most consequential failures of the next year, across all three products, to be less about a dramatic single mistake and more about the slow erosion of anyone actually checking the log — a pattern with decades of precedent in every other domain that's paired automation with human review.

The products that survive the next year won't be the ones with the most features at launch. They'll be the ones whose safety mechanisms were still holding up once nobody was watching the launch-week headlines anymore.

The regulated sectors will decide the real ceiling

OpenAI's admin-enabled beta for Enterprise, Edu, and Healthcare — deliberately slower than its consumer rollout — is a bet that the biggest durable revenue in this category sits in institutions that move carefully. If that bet pays off, dots' Agent 365 integration with Microsoft becomes the template every competitor has to match to be taken seriously by a hospital system's compliance team or a school district's IT department. If it doesn't — if enterprises decide the risk isn't worth it regardless of how careful the rollout looks — always-on agents may end up confined to the consumer and SMB tiers where Muse is already strongest, a very different ceiling on the category's ultimate size.

What won't matter as much as it seems to right now

The cartoon avatars, the dot.com stunt, the app-chart rankings — all genuinely fun to write about, and none of it will be a meaningful factor in which product wins durable enterprise trust a year out. Branding solves a first-use anxiety problem; it doesn't solve the underlying question of whether an agent with standing access to your tools behaves safely at scale, month after month, across use cases nobody demoed at launch.

Where this leaves the skills ecosystem

None of this changes the Claude Code Agent Skills calculus much either way. Skills solve a different problem — narrow, inspectable, human-invoked capability rather than standing autonomous access — and that distinction holds regardless of how the always-on-agent race shakes out. If anything, a year of real-world incidents and safety-documentation gaps across dots, Grok Bot, and Muse would make the case for composable, versioned, git-inspectable capabilities stronger, not weaker — a reminder that the most durable trust in this industry has always come from being able to see exactly what you installed, not from how friendly its avatar looks.

This series set out to cover what launched in three weeks in September 2026. What happens over the following year — the pricing, the incidents, the enterprise adoption curves, the safety documentation that either arrives or doesn't — is the more important story, and one worth returning to as it unfolds.

Part 30 of 30 in the ChatGPT Dots series. Previously: Enterprise, Edu, and Healthcare: The Slow Lane for Dots. This is the final article in the series. Browse more builder insights or explore AI skills at aiskill.market.