Gemini agent, Dots, Muse, Copilot and Cue: how far can we entrust work to AI?

Gather internal information for next week’s customer meeting, build a briefing, update it when things change, and ask before sending it. That request needs more than one well-written response. It needs a system that remembers the responsibility, reaches the relevant apps and stops at the right boundaries.

Google’s Gemini agent announcement on October 8, 2026 brings planning, execution, memory and administration under one entry point. OpenAI Dots, Meta Muse, Microsoft Copilot Autopilot and Manus Cue also address ongoing delegated work. The question here is how far the systems needed to entrust work to AI have developed.

Comparison date: October 9, 2026. This is a review of official announcements, help pages and administration documentation, not a measured five-service task trial. “Unconfirmed” means the inspected material does not establish the capability, not that it is unsupported.

Google announcement of October 8

Key takeaways

1. All five address work that continues after an app closes. But new Gemini capabilities include early access, Copilot Autopilot is in private preview, and Cue remains invite-only early access. Announcement scope is not general availability.
2. Delegation depends on execution, memory and connections, but also on whose authority is used, where approval is required and what is recorded. Cue’s separate email addresses and wallets do not by themselves establish enterprise governance.
3. A subscription does not necessarily mean unlimited long-running execution. Check billing units, allowances, stopping behavior and Japanese availability against the actual workflow.

Separate a deliverable from an ongoing responsibility

Producing a deck has a completion point. Keeping that deck current with customer feedback until a meeting is an ongoing responsibility. It adds scheduled wake-ups and reactions to events such as incoming mail. “Always on” alone does not explain what starts work or how far it proceeds.

Gemini work and integration overview

Dots tasks and local access

Muse for Small Business use cases and connections

Copilot Home, Code and Autopilot announcement

Cue here is the standalone app announced by Manus on September 28. The official blog links to cue.im; app stores name BUTTERFLY EFFECT PTE. LTD. as its developer. It is distinct from similarly named Cue OS and voice services.

Manus announcement of Cue

Announcement date in the Manus blog index

From a delegated responsibility to a result

Conceptual example: keep a customer briefing current

Trigger and context

A scheduled check or new message leads to memory and authorized source material.

Execution and boundaries

Research and update the briefing; check recipients and sharing scope; return to a person when required.

Record and continuation

Record results and actions, then resume when needed. Stop when budget or authorization is exhausted.

Separate announcements from access

Google’s announcement is broad, but the current product FAQ explicitly places multi-step background delegation, mobile and desktop access, and third-party model choice in early access. The announcement describes Gemini and Claude selection, while the FAQ says Google models are available now and additional models are coming. The full new experience cannot be treated as available today under every contract.

Consult the current Gemini FAQ for availability

Dots plans and gradual rollout

Muse age and region eligibility

Availability stages for Copilot capabilities

Current Cue iOS listing and invitation requirement

Cue Android listing, updated October 8

Where work runs after the laptop closes

Gemini execution and triggering design

Dots cloud and local administration conditions

Muse background-work design

Autopilot cloud execution

“Continues after you close your PC” does not mean applications on that PC continue running. A job using a VPN-only site or local files depends on where execution happens. Test cloud sign-in requirements and stopping behavior when a local connection disappears.

Cue-specific background execution and automation

Manus Studio Remote Control was announced alongside Cue, but is a separate capability. The existence of a Cue desktop app does not establish the extent of direct control over the user’s computer.

What memory and app connections carry forward

Gemini describes four memory categories: active context, business knowledge, procedures and past actions. Dots receives ChatGPT memory and creates its own. Muse retains context across conversations, with readable and editable Memory files described in its design article. Autopilot has its own memory and business context through Microsoft IQ.

Memory availability does not establish complete control over updating and deleting it. Dots help says disconnecting an app does not erase information already obtained, and deleting the dot is required to remove its own saved memories. Plan what remains after a responsibility ends.

Gemini memory and integrations

Dots memory and disconnection behavior

Muse Memory files

Muse business connectors

Connector count is not a count of completed workflows. Read access, write access and event delivery enable different work. Reading a Slack message does not imply permission to forward it by external email. MCP or A2A connectivity also does not establish portability of memory or approval policies between services.

Cue describes per-agent conversations and memory, with Memory and Dream retaining preferences and decisions. API documentation also identifies a persistent main-task conversation for agents created in Cue. The detailed mechanisms for selecting or sharing memory remain unconfirmed.

Cue memory and connected apps

Persistent conversations in the Manus API, including Cue agents

Deletion has a broader boundary too. Official help says Cue and Manus share an account and data; separate Cue-only account deletion is unsupported. Some information may be retained for legal or accounting purposes. This is separate from whether individual memories can be edited.

Account-deletion scope across Cue and Manus

Multiple agents and independent identity

Gemini distinguishes temporary subagents for parallel or sequential work from persistent coworker agents with a continuing role. Coworkers receive dedicated email, storage and identity. Splitting computation is different from registering an accountable organizational role.

Specialist dots with organizational identities have been announced, but in enterprise pilots. They should not be folded into the standard personal-dot feature set. Teams of primary dots working together are also presented as a future direction, distinct from a current dot delegating tasks to Work or Codex.

Muse documents concurrent tasks and an independent Sentinel that checks actions. That does not by itself establish a general enterprise framework for composing arbitrary specialist teams. Autopilot has its own identity and workspace, but the inspected sources do not establish an equivalent scope of dynamic subagent orchestration.

Gemini subagents and coworkers

Specialist-dot pilots and future collaboration

Separation of Muse and Sentinel

Autopilot identity and workspace

Cue emphasizes separate email, phone, wallet and computer resources for each agent, with work handed between roles in a group chat. Current developer descriptions qualify those resources by availability. They do not establish that every user receives all four or that these identities integrate with enterprise identity governance.

Divide a job among three Cue agents

Read the announced venue-planning example through its handoffs

Research agent

Collect venue information, then hand it to the comparison agent.

Comparison agent

Build a shortlist, then hand it to the drafting agent.

Drafting agent

Prepare a deck; the person makes the final decision.

Cue identities and group collaboration

Questions, approval and access are different

Permission to read is separate from approval to send

Three checks when sending a customer briefing

Is context sufficient?

Ask which customer, meeting date and purpose. This clarifies the task.

Is access authorized?

Reach the relevant file and mail service within granted access.

Is this action approved?

Present recipients and attachment contents, then obtain required approval.

Dots Custom Rules and Auto-review

Muse action approval and egress controls

Autopilot authority and approval

Proactive research is also different from proactive sending. OpenAI says proactive research uses read-only tools and those research tasks cannot directly message people or modify connected apps. Muse for Small Business also requires approval for sending, publishing and spending. The continuing scope of a previous approval still needs checking against settings and the particular job.

Muse sending, publishing and spending conditions

Cue permissions and payment confirmation

Cue’s launch describes spending within a budget, while current store text also specifies payment confirmation. That does not establish unrestricted purchasing without review. Clarification, pre-send checks and revocation of contact or wallet permissions still require workflow-level testing.

Check activity records and integration exceptions

Google announces agent-attributed logs, sandboxes and Agent Gateway controls. However, the documented custom A2A registration route sends traffic outside Agent Gateway, so its policies do not apply there. Console Model Armor settings do not automatically protect those agents either; developers must configure protection in their application code.

That is not a claim that Google has no protections. It is an integration boundary where responsibility and enforcement change, even if the agent is launched from the same interface. Check who retains logs and where authorization and data egress are enforced after adding an external agent.

Security conditions for custom A2A registration

Muse describes activity and permission views; Dots shows in-progress, scheduled and completed work; Autopilot describes permissions, audit and governance. A user-facing activity history does not establish retention, tamper resistance or export suitable for an organization’s audit needs. These three properties were not verified under matched conditions across the five services.

Privacy features also have rollout stages. Do not conflate Muse Secure VM with the subsequently planned Muse Confidential VM. Dedicated execution, credential isolation and encryption that prevents provider access are separate properties to verify.

Current Muse VM and planned Confidential VM

Muse activity and permission views

Cue-related agents can expose subtasks, messages and stopping controls through the API. These aid inspection but do not establish tamper-resistant audit logs, administrator exports or retention guarantees. Those Cue-specific conditions remain unconfirmed.

Agent subtasks, messages and stopping controls

Separate seat prices from the cost of continued work

Prices above are official US-dollar figures, not yen invoices or tax-inclusive total costs. Business Standard and Premium are different; the minimum is two paid seats, which may mix seat types. Dots’ extended introductory allowance should not be read as unlimited normal use.

Google subscriptions and consumption billing

ChatGPT Pro pricing

Business seat types and pricing

Muse allowances and pricing

Copilot USL and UBB

Muse tokens, Copilot Credits and model API tokens cannot be converted into a common unit just from their names. Compare cost per completed task with model choice, input, runtime, retries and external-service charges accounted for. Long-running cloud execution does not guarantee inexpensive continuous work.

Google’s announcement uses hard-limit language, but its administration documentation says enforcement may take several minutes and charges can exceed the limit. Setting a cap is not a guarantee of zero overshoot. A workflow also needs a way to establish what completed before a budget-triggered stop, before resuming.

Spend-limit enforcement delay and possible overshoot

Cue’s free-access description and listed in-app purchases are different evidence. A price list without billing periods and eligibility cannot establish monthly cost. Verify phone numbers, call allowances, computers and actual purchases in the relevant checkout terms.

Cue free Plus and in-app purchase listings

Official membership help on Cue phone charges and weekly allowance

Check access from Japan

Japan is a supported ChatGPT country, and Dots is offered on eligible regional Pro and Business Premium plans. Japan therefore meets the stated regional condition, while plan eligibility, rollout and organizational settings still apply. Do not extend individual features such as the US texting beta to Japan.

Supported ChatGPT countries

Dots regional and feature-specific conditions

The current Muse business page lists US and Canadian adults, not Japan. Autopilot is in private preview; the availability of Microsoft 365 subscriptions in Japan does not establish access to it. Likewise, Japanese Gemini customers or a Japan region do not establish general availability of every new capability in Japan.

Google’s locations documentation describes the Japan region as GA with an allowlist. A domestic storage arrangement is separate from ensuring that model processing, integrations and logs all remain in Japan. Check Japanese UI, output quality, domestic storage and inference, and eligibility for each new feature separately.

Google locations and access conditions

Cue is listed in Japan’s App Store with 22 languages including Japanese, requiring iOS 17 or later on iPhone and an invitation. Store availability does not establish Japanese phone numbers, SMS, wallets or merchant coverage; these domestic conditions remain unconfirmed.

Cue Japan listing, OS and language requirements

Test delegation on one bounded workflow

The supporting systems now extend to execution environments, persistent memory, app access and action controls. But personal delegation, an organizational teammate and a developer-built agent platform have different premises. This comparison does not establish a universal winner for every organization.

For a customer-briefing trial, start with permission to read internal sources and update a draft, then return to a person before external sending. Test duplicate events, expired sign-ins, mid-task budget exhaustion and a change of owner. Evidence about duplicate sends, stale-memory updates and traceable stopping reasons makes it possible to expand responsibility deliberately.

Record more than task counts: deliverable correctness, time spent reviewing, and cost including corrections and reruns. Moving from “Can it create a document?” to “Under what conditions can it own this responsibility?” makes the differences more useful.

Sources and scope

This comparison uses public sources checked on October 9, 2026. It is not an account-level feature test, recovery test, audit certification or price quote. Because rollout descriptions differ, announcement scope is paired with current FAQ and administration conditions. The figures are LATENT’s conceptual comparisons.

Cue was also searched on X, identifying Manus’s introduction and phone-feature posts. Direct retrieval returned 403, so the videos and full threads were not verified. User anecdotes are not treated as behavior guarantees; specifications rely on official blogs, developer store descriptions, API documentation and help.

X: official Manus Cue introduction (search-identified; video unverified)

X: official Manus phone-feature post (full text unverified)