Your personal agent should leave you with less to do

A Number That Matters

40 minutes

That is how long The Verge’s Allison Johnson says she had been using Meta’s Muse before she had supplied her name, address, payment details, a backyard photo and access to her home-security account.

Muse moved neglected chores forward. The permissions accumulated just as quickly. This is one reviewer’s experience, not a typical onboarding time or a claim that the model could see every credential.

The Verge — Muse hands-on, September 23

This Week’s Thesis

Our read: personal agents should be judged by the responsibilities they remove, not the activity they generate.

Meta’s Muse, OpenAI’s Dots and Microsoft’s Copilot Autopilot promise an assistant that stays with the work. OpenClaw and Hermes let operators build that continuity on infrastructure they choose. The meaningful difference is less “agent versus harness” than who assembles, controls and repairs the system.

Early field reports describe useful work alongside a less glamorous problem: people still have to find approvals, recover stalled tasks and reconnect context. An agent can complete more work while leaving its owner with more to manage.

What Supports the Thesis

Different products, different places to delegate

Muse starts with everyday life. Meta describes a cloud computer with a browser, files, tools and scheduled work. Its design account emphasizes a continuing conversation, goals and selective proactive messages. In Johnson’s hands-on test, Muse submitted requests for tree-removal quotes and helped purchase supplies. But downgrading a security plan still required her to make the phone call. Useful progress did not always mean a finished chore.

Meta — Muse product design · The Verge — completed work and unfinished chores

Dots starts with an ongoing relationship inside ChatGPT. OpenAI says a Dot has its own cloud computer, carries context across supported channels and learns from feedback. The September 29 launch began a rollout to eligible Pro and Business Premium users, with an admin-enabled Enterprise beta. The first Dot is included in eligible paid plans; deeper work and delegated Codex or ChatGPT Work tasks have separate allowances or limits. “Always on” is not unlimited execution.

One important boundary: OpenAI says unsolicited proactive research uses read-only tools. An authorized background task, by contrast, may take actions under its permission rules.

OpenAI — Dots launch, availability and safeguards

Autopilot starts with the employer’s work environment. Microsoft’s September 25 announcement describes the product formerly called Scout as cloud-hosted, with its own identity, memory, computer and workspace inside the organization’s tenant. It is intended to work through Teams, Outlook, channels and documents. Microsoft announced an expansion of its private preview—not general availability—and usage-based billing for Autopilot work.

Our read: this is delegation attached to an employer’s systems and authority, not simply a personal chatbot with more features. Its proposed supplier-review workflow remains a vendor example, not an independently established outcome.

Microsoft — Copilot Autopilot announcement

Open harnesses make a different bargain

A harness connects a model to tools, memory, permissions and recurring work. Packaged agents need that machinery too; they conceal more of its assembly.

OpenClaw documents an operator-run gateway and Markdown memory files. Hermes documents persistent memory, reusable skills, scheduled work and a choice of model providers and execution environments. Neither is “just a chatbot.” Nor does their flexibility establish superior reliability.

OpenClaw — platform documentation · Hermes — platform documentation

Our read: managed products trade some operator choice for a more assembled experience. Open harnesses leave more room for custom workflows, but someone still owns updates, backups and recovery. Operator control is not automatic privacy: information may still travel to external models and connected services.

For a buyer, the question is not whether the system is technically open. It is which responsibilities the supplier actually takes off the buyer’s hands.

The community is testing the handoffs

Weak signals—not failure-rate estimates: original posts on the OpenAI Developer Community describe useful initiative and the work it can leave behind. These are self-selected practitioner reports, not independent comparative testing.

• A useful catch. A forum moderator reports that a Dot noticed a delayed flight in connected email and suggested reconsidering a connection. After a second delay, it alerted him again; he says rebooking became necessary. This is a first-hand self-report, not independent testing or proof that every alert is well judged. OpenAI Developer Community — flight-delay report

• Approval hunting. Two participants describe permission requests that do not surface where they are actively chatting. One says they have to open individual sidebar tasks to discover what is waiting for approval. A pause the person cannot readily find creates another monitoring job. OpenAI Developer Community — permission-request thread

• Useful retrieval, incomplete delivery. A participant praises a Dot for finding old conversations but says it does not supply a direct link. Their workaround is to copy a distinctive phrase into ChatGPT search and find the conversation manually. They also report uncertain persistence of files on the Dot’s computer. Those are their observations, not a verified platform-wide retention policy. OpenAI Developer Community — workstation and conversation-retrieval report

These reports suggest specific tests: does a useful discovery arrive with an actionable next step? Does a blocked task bring its blocker to the person? Can tomorrow’s work pick up what today’s work produced?

They do not establish that Dots is unusually unreliable. We do not have equivalent field samples for every product in this comparison.

Remembering you is not the same as letting you manage the memory

Meta’s design documentation says Muse’s memory files can be read and edited directly. OpenAI’s Dots FAQ says users currently cannot view, delete or directly modify individual Dot memories. Deleting the Dot deletes its own context, but separately stored files, conversations and ChatGPT memories have separate controls.

Meta — memory-file controls · OpenAI — Dots privacy, security and safety FAQ

OpenClaw’s documentation describes importing Markdown memory from Hermes, Codex and Claude Code. That is a concrete portability mechanism, not proof that moving notes reproduces an agent’s understanding or behavior.

OpenClaw — memory and import boundaries

Our read: as these relationships accumulate useful context, correction and exit become practical buying criteria. The question is not only “Will it remember?” It is “Can I inspect what it retained, correct it and leave without starting from nothing?”

What Challenges the Thesis

The strongest counter-case is that packaged agents need not be perfect to be valuable. Removing most of an annoying chore can be enough.

Johnson’s Muse test supports that case. The agent helped move work she had postponed; even an incomplete task could lower the barrier to finishing it herself. A useful flight warning need not eliminate every approval to earn its place.

The Verge — partial completion with practical value · OpenAI Developer Community — unsolicited warning

Launch-period bugs may also be temporary. Open harnesses are not a repair-free alternative. We still land on attention returned as the test—not complete autonomy as the requirement. A good assistant can ask for help. It should make that help easy to provide.

What to Do With This

Give one agent one recurring, low-risk responsibility for two weeks: watch a project’s approved sources, prepare a weekly status draft or flag a defined class of changes. Start read-only; do not connect everything merely because the setup screen offers it.

Track useful results, missed obligations and time spent on setup, approvals, checking and repair. Compare that with the old way of doing the same job. Keep money spent and consequential errors visible alongside the time record.

The proposed measure is net attention returned: the effort the responsibility used to consume, less the effort required to operate the agent. This is an evaluation frame, not a reported benchmark. Count partial completion honestly—and do not call a larger pile of drafts a smaller workload.

How Sure Are We?

Known: the cited public documents describe product capabilities, memory controls and rollout stages. Our read: responsibility removed is a better selection criterion than agent activity. Weak signal: the community reports identify concrete friction worth testing, not its prevalence.

We have not run a controlled comparison or a hands-on comparative test. Product documentation, one Muse reviewer’s task account and a small Dots forum sample are an uneven evidence base. A forum contribution from a contributor involved in this publication is not independent corroboration. Autopilot remains preview evidence here. We are not declaring a winner or treating the absence of comparable reports as proof of quality.

What we expect next—conditional, not established:

• Over the next 6–12 months, managed agents could make recurring delegation easier to begin, while open harnesses remain useful for bespoke work. This read would weaken if packaged products retain little repeat use or offer equivalent operator choice and portability.
• Memory correction, clear blocked-work notices and dependable recovery could matter more in repeat-use decisions than another long feature list. Continued delegation of the same responsibilities with fewer interventions would support this read. It would weaken if capability breadth consistently delivers greater measured savings despite that friction.

The evidence that would most change our view is sustained use: people delegating the same responsibilities week after week, spending less time supervising them and still achieving acceptable outcomes.