The Brief That Outlasts the Chat

Most of us learned to work with AI in a particular rhythm. You open a conversation, explain the situation, get something useful and close the tab. Tomorrow you explain it again.

Products like Dot, Grok Bot, Muse and Claude’s Cowork are built around a different rhythm. The work can continue after the conversation ends. That sounds like one feature, but it is several, and each one asks a different question of you.

This piece pulls them apart. Then it follows one fictional project through a few weeks, so the differences stop being abstract. It ends with a one-page brief you can adapt for your own ongoing work.

---

The word “persistent” is doing too much

When a product says an agent is persistent, it can mean five separate things. They are easy to blur together.

1. Remembered context. Something from earlier is available later: a preference, a note, a fact about your project.
2. Source access. The agent can look at specific places, such as a feed, a folder, a store or a browser.
3. A stored responsibility. The job itself is written down somewhere: what to watch, what counts, what to return.
4. A trigger. Something causes a run: a clock, a supported event, or you.
5. A delivered result. Something useful reaches you in a place you will see it.

An agent can have any of these without the others. A tool might remember your tastes but never run unless you open it. Another might run nightly but remember nothing you told it last week. A third might read a wonderful source and deliver the result into a place you never check.

None of this requires believing the agent “knows” you or “cares” about the project. It requires only that the pieces exist and that someone decided what each one holds. That someone is you.

Scroll sideways to compare columns.

Piece	The plain question	Who decides
Remembered context	What should still be true next month?	You, by saying it and revisiting it
Source access	Where may it look?	You, by connecting or declining
Stored responsibility	What is the job, and what is not?	You, in writing
Trigger	What causes a run?	You, within the product’s controls
Delivered result	What shows up, and where?	You, by shaping the output

These are choices the person can help shape within the service’s available controls and permissions. The agent contributes work. The shape of the job remains an authored thing, and that is the exciting part: you can edit it as the project changes, instead of re-prompting from scratch.

---

A designed example: the harbor wall

Everything in this section is a designed scenario. It is not an observation of a real user, product run or result. It is built to show decisions, not performance.

Meet Ines, an invented exhibition designer. She has been asked to build a small show for a community maritime museum: one wall, three meters wide, about the working life of a harbor ferry. She has eight weeks and a modest loan budget. She wants an agent to hold an ongoing watch on the question “what could belong on this wall?” while she gets on with the rest of her life.

Her stored responsibility, in plain words:

“Look for objects, documents and images related to the ferry that could be borrowed or reproduced. Return only what is new and relevant. Tell me why it matters for the wall.”

Four things then happen over the weeks.

The initial preference

Ines tells the agent early on that she prefers primary material (logs, photographs, repair records) over blog roundups and listicles. She also likes objects with visible wear, because wear tells a story at a glance.

This belongs in remembered context. It should survive for the length of the project, and probably into the next one. It is a standing taste, not a fact about this week.

The temporary constraint

Midway through, Ines travels for a conference and will be unreachable until the twelfth. She writes: “Don’t send anything that needs a decision from me before I’m back. Collect it instead.”

This is a constraint with an end date, so it should expire on that date. If it quietly stays forever, the agent keeps holding back results long after Ines is home. That is the failure mode of memory without expiry: yesterday’s exception hardens into permanent policy.

The surprising source

In week four, a municipal archive posts digitized logbooks kept by the ferry’s engineers. Ines had not thought to look for them, and neither had her list of sources. Suppose the agent is allowed to read that archive and flags the new collection.

Here the agent is doing something useful that has nothing to do with intelligence in the grand sense. It is checking a source she granted, noticing a change and describing why it fits her stored brief. The surprise is in the world, not in the machine.

The changed priority

Ines reads a few pages and realizes the margins are full of engineers’ handwritten complaints about the weather, the engine and each other. The exhibition has been an object show. Now she wants it to be a show about voices. Three of her four candidate objects no longer fit.

This is the moment that separates a stored responsibility from a stored list. A list would keep returning objects. A responsibility she can revise lets her say: “Priority change. Handwriting and marginalia first. Objects only if they connect to a named person.”

The useful response

Two weeks later the agent returns a short note:

“One thing changed since the last note. A second volume of engineers’ logs was added to the archive. It covers the winter your brief mentions. Three pages include marginal notes about the engine room lamp; a lamp matching that description is listed in a regional collection. Nothing else in your sources met the new voices-first standard this period.”

Look at the last sentence. Much of the value is in what it leaves out. A week with no meaningful change should produce a short quiet note, or no note, rather than a padded one. An excellent ongoing agent can stay quiet. Activity is not the goal. The quiet rule applies when the check succeeded and found nothing relevant; a blocked or failed check needs a different report.

---

What deserves attention, what expires, what stays fresh

The Harbor Wall example gives three kinds of information, and each one needs different treatment.

• Remember. Durable taste and project facts. Ines’s preference for primary material and visible wear.
• Let expire. Anything tied to a moment: the travel constraint, a budget that applies until the loan deadline, a “focus on this archive this week” instruction.
• Ask again. Anything where the right answer depends on the situation when it arrives. Whether a particular lamp is worth a loan request is not a stored rule. It is a fresh judgment, and the agent’s job is to bring it to Ines at the right moment with the evidence.

That third category is the one people skip. The temptation is to turn every judgment into a rule, so the agent can handle it alone. But some decisions should return to you each time, because the context you hold (a curator’s reaction, a gut feeling about the wall) is not written anywhere.

---

How the products divide the pieces

The products differ in how they handle the five pieces. A brief comparison helps, but it is not a ranking. These are documented intentions as of 15 October 2026, not tested results.

Dot. OpenAI’s documentation describes dots as coordinating ongoing responsibilities, delegated tasks, schedules and supported events. Memory draws on selected context, relevant ChatGPT memory and separate saved notes, not every full conversation transcript. Local tasks keep their connected-computer requirements. A responsibility does not have to be a fixed-time schedule; when timing matters, use the documented controls. That is a useful reminder that “ongoing” and “scheduled” are different.

Grok Bot. xAI introduced Grok Bot on August 11, 2026 as a persistent teammate in beta, with cloud computers, learned routines and coordinated Bots. The internal examples there are vendor-selected. On August 29, xAI described an X connector for timeline, mentions and search, with a plugin that adds trends and bookmarks. That is a source surface. It does not show that the social material is representative or that discovery works well.

Muse. Meta introduced Muse on September 8 (updated September 30). It describes a dedicated virtual machine and browser, interaction through an app and WhatsApp, and goals and remembered preferences. Meta describes approval steps and protections, and no independent test was run for this feature. A confidential VM is described as coming later this year, so it is not part of what exists now. The launch is a US rollout. A small-business expansion (September 29, updated October 8) lists US and Canada, store, brand and business connectors, proactive suggestions and Ideas, with approval before publishing, sending or spending described by the vendor.

Claude Cowork. Anthropic’s September 16 article describes merging task and chat modes and editable Docs, Slides and Design inside conversations. For timing, the current scheduling help page says new recurring work runs remotely. Legacy Pro and Max computer schedules created before October 6 stay local, and any step touching local files or apps still needs that computer. Do not assume a laptop can always close.

OpenClaw. The OpenClaw documentation on cron jobs is currently an Automations overview. It describes persisted jobs, wakes and output destinations, with separate pages for execution, delivery and reconciliation. It establishes that those are distinct concerns. It is not a setup recipe, and this article does not offer one.

What can be concluded from all this? The documented designs treat memory, access, scheduling and delivery as separable. What cannot be concluded is how reliably any of them work in a particular account, how good the discovery is, or whether results justify the effort. No comparative tests, multi-week runs or customer outcomes were supplied for this feature, so none are claimed.

---

Dated note

Sources read 15 October 2026. Product access, plans, regions and features were not tested in any account and may change. Check the linked pages for current terms.

---

The responsibility brief (adapt this)

A responsibility is easier to improve when it fits on one page. Here is a template, with Ines’s answers beside it as a designed illustration.

1. The question. What is this job really about? Ines: “What could belong on a three-meter wall about the ferry’s working life?”

2. Standing taste (remember). What stays true across the whole project? Ines: Primary material over roundups. Visible wear. Named people over anonymous objects.

3. Sources allowed. Where may it look, and where not? Ines: The municipal archive, two regional collections, her own saved notes. Nothing behind a login she hasn’t granted.

4. Trigger. What starts a run? Ines: Check weekly, plus any new item in the archive, if the product’s controls support that. Otherwise she runs it herself.

5. Temporary rules (let expire). What applies only for a while, and when does it end? Ines: “Collect, don’t ask” until the twelfth. Loan budget cap until the deadline.

6. Fresh judgments (ask again). What must come back to me every time? Ines: Any loan request. Any change to the wall’s theme. Anything involving contacting a person.

7. The quiet rule. What should happen when nothing meaningful changed? Ines: One line saying so, or nothing.

8. Delivery. What does a good result look like, and where does it land? Ines: Three sentences, one link, one reason it matters for the wall. In the inbox she checks on Friday.

9. Review date. When do I reread this brief and edit it? Ines: Every second Friday, or whenever the theme changes.

The review date is the quiet hero. A brief with no review date drifts out of sync with the project, and the agent keeps faithfully serving a version of you that no longer exists.

---

The next decision

Writing the brief exposes something the chat window lets you avoid: your own preferences, stated plainly, with dates attached. Ines discovered mid-project that she wanted a different show. The agent did not cause that. What helped was that her brief was explicit enough to revise.

So the interesting decision is not which product to try first. It is which part of your taste you are willing to write down and keep honest. Start with one responsibility small enough to review in five minutes, and put a review date on it before anything else. Then see which of its lines you rewrite first, because that line is probably the one you did not yet know you believed.