Personal AI agents: where we think each earns its place

On this page

Grok Bot can read X. A dot can start a Codex task. Instinct offers to order groceries. Those are three quite different reasons to delegate work, even though all three products appear in the same personal-agent conversation.

Our first comparison of dots, Muse, Grok Bot and Instinct examined the launch propositions. We followed it into setup instructions, connected apps, billing pages and terms, then widened the investigation to Hark, Gemini Spark, Lindy, Manus and managed OpenClaw. We asked what gives each offer a concrete advantage, what the documents substantiate and what would still need proving through use.

We came away with a more useful opinion than an overall ranking. Grok has a clear case for research on X; dots is especially interesting as a coordinator of OpenAI's working tools; Instinct makes personal purchasing and errands central to its offer. Hark's attraction is a persistent assistant across Google and Microsoft accounts. Muse's business connections, Spark's Google integration and Lindy's follow-up routines give them more specific roles.

We haven't run a comparative product trial. The work behind this article is a cross-check of primary sources on 9 October 2026. Our judgements concern the offers those sources establish, rather than speed, accuracy or successful purchases we haven't measured.

The advantages we found worth taking seriously

We used the same question across the category: what does this product make meaningfully easier to hand over? This is our assessment of the documented propositions.

Agent or arrangement	Where we see its strongest case	The qualification that matters to our view
Grok Bot	Following public X conversation and carrying research into persistent work	Public reading, private-account access and the @bot shortcut are different features
OpenAI dots	Coordinating continuing projects, including Codex and ChatGPT Work tasks	Local work needs enabled computer access; downstream tasks retain their own usage limits
Instinct	Personal purchases and errands requested through texts or calls	We haven't completed a purchase or established Australian availability
Hark	One continuing assistant for web errands and Google/Microsoft work	Usage tiers, top-ups and standing transaction permission affect the arrangement
Gemini Spark	Recurring jobs across Gmail, Calendar and Drive	Personal account eligibility differs from business access
Meta Muse	Operating across a small business's commerce, accounts and social tools	Current announcement names US and Canada
Lindy	Following unanswered messages and administrative routines	Draft-only flows and channel-specific guardrails offer different controls
Manus	Carrying a research or production project through a workspace	Persistent Cloud Computers are a separate purchase; Cue is a distinct early-access app
Managed OpenClaw	Configurable agents without operating all the hosting infrastructure	Runtime operation, model access and useful workflow configuration remain distinct

The table isn't a claim that only one agent can perform each job. It names the advantage we find most persuasive in each offer, and the evidence that keeps our enthusiasm proportionate.

Grok Bot: X is a concrete research advantage

For the kind of market research we publish, X is a consequential source. Launch claims, implementation questions and user objections can sit in the same thread. We think an agent with a direct route into that material has a useful starting position.

xAI's computer and apps guide documents public X search and reading without an account connection. The X plugin adds signed-in material such as DMs, bookmarks and the home timeline. We describe the Bot capability specifically rather than assuming it includes every developer X API endpoint.

We can see a worthwhile assignment here: follow a release, gather the supporting posts and preserve a sourced account of what changed. That is an illustrative use we would investigate, not a result we've obtained. The advantage lies in the material the agent can reach and retain, not an assumption that social content is accurate.

There is a local qualification. The @bot handoff can turn a tagged post into a Bot task, but currently excludes Australian accounts. That doesn't erase public reading. It does stop us describing every X integration as available here.

Our favourable view of the research proposition also survives a less favourable view of its price clarity. Our Grok Bot billing investigation found non-stacking linked allowances and unresolved instructions about mixed plans. We like the access advantage. We still need a measured research job to assess its value.

Dots: the interesting role is coordinating the work

The detail we find most compelling about dots is its relationship with Codex and ChatGPT Work. A continuing assistant can hold the objective while specialist tools do parts of the work. That is a credible answer to a problem familiar to anyone juggling several AI tasks: somebody still has to keep the project together.

OpenAI's launch announcement explicitly describes dots starting or managing tasks in Codex and ChatGPT Work. Those tasks still count towards the respective usage limits. Its setup guide says enabled local access lets a dot create local Work or Codex tasks; cloud Codex tasks require an environment prepared in Codex.

We think that makes dots especially interesting for software and other projects already spread across OpenAI's tools. An illustrative flow is a continuing conversation about a product change, a Codex implementation task and working material prepared elsewhere. The coordinating role is the point we would evaluate: does the dot preserve the objective and bring the parts back together?

That isn't a claim of flawless project management. Access, handoffs and review still determine the result. Current eligibility names Pro, Business Premium and an admin-enabled Enterprise beta, with a gradual rollout. Local computer access starts turned off. The proposition is stronger when described through those actual working relationships than as an assistant that can do everything.

Instinct: purchases make the proposition tangible

Instinct's homepage names groceries, gifts, disputed bills and appointments. We find that persuasive because the requested result exists outside the conversation. Groceries arrive, a charge changes or an appointment is made. Success can be checked against the world.

Its text-and-call interface also fits those jobs. An errand can be requested during the day without turning it into a project inside another dashboard. We see the appeal in combining that familiar channel with the ability to use a phone and computer.

We followed the purchase claim into Instinct's terms. They explicitly address purchases on connected services, including passing relevant information to the merchant or payment provider. They say the customer makes the purchase, with the merchant's prices, cancellation and refund rules applying; Instinct facilitates rather than processes the payment.

That is more substantive than a shopping recommendation. It describes an assistant acting in a transaction. We think this is one of the clearest personal-agent propositions in the group, while stopping short of claiming a completed or reliable purchase from our own experience.

Our unanswered questions are practical: current price, Australian access and performance when an order needs a substitution or refund. The terms also grant a personal-use licence. We haven't treated that as permission to deploy it across a company's purchasing. The commercial potential interests us precisely because real transactions introduce obligations that a chat answer doesn't carry.

Hark: a continuing assistant across two major work ecosystems

Hark broadens the errands proposition. Its Pro introduction brings memory and a cloud computer into a persistent conversation; its operating guide documents schedules and Google Workspace and Microsoft 365 connections.

We think the cross-ecosystem arrangement deserves attention. Work is often split between accounts and websites. A continuing assistant that can use both suites has a clearer purpose than one confined to the conversation where the request began. We haven't measured how well Hark carries that context through a real errand.

Its commercial structure is unusually concrete in the material we checked: the feature set exists on a capped free offer; Pro² displays $20/month for twice the usage and Pro³ $100/month for ten times the usage. Optional automatic top-ups add further capacity. Those displayed amounts aren't Australian checkout quotes.

The terms, revised 1 October, say purchases and communications require approval unless instructed otherwise. Standing permission can replace individual confirmation. That changes the assessment: Hark is offering continuing action under an agreed arrangement, as well as continuing conversation. We like the clarity of that proposition; we still want observations of how it handles a changed instruction.

Spark and Lindy: the recurring obligation is the useful part

We find Gemini Spark's strongest case in its combination of Google apps, reusable skills and schedules. A weekly review drawing from Gmail, Calendar and Drive is an intelligible ongoing job. Google's connections give that proposition a specific foundation.

Its operational help distinguishes remote-browser work from local Chrome work that needs an awake device. Personal setup also requires an eligible adult account, Pro or Ultra and Keep Activity. That keeps our view tied to a defined execution route, rather than a universal promise of background work.

Lindy's follow-up routine gives an even narrower example: watch sent messages and prepare a bump when a reply hasn't arrived. We like it because the assistant assumes an obligation people easily forget. Built-in drafts retain the human sending step.

The qualification is channel-specific authority. Lindy's integration guardrails cover writes in shared Slack threads, not web chat, Slack DMs, iMessage or SMS. The email and calendar investigation follows that distinction and the native Google/Microsoft alternatives in detail. Our opinion is that follow-through gives these offers their purpose; another drafting box would be a weaker case.

Muse and Manus: operating a business and producing a project

Meta's Muse for Small Business announcement names Shopify, QuickBooks and business social connections. We think that gives Muse a coherent commercial direction. Those systems contain the orders, accounts and customer interactions an operator already manages.

The significance is access to the business's working systems. A general assistant can discuss an order; an operator needs to work with its actual record. We find Muse's named connections more informative than the broad label “personal agent”. Current announced availability is US and Canada, so our Australian assessment remains about the proposition rather than available account access.

Manus 2.0 presents another arrangement: a project workspace with connected-service events and automations. We see the stronger case in an evolving deliverable, such as a research report or document, where the work needs a place to accumulate.

The announcement separately describes purchased persistent Cloud Computers and Cue, its early-access personal-agent app with invite codes. We haven't combined those into one subscription promise. Our view of Manus is strongest where a project workspace is the product being evaluated; current Australian entitlement remains unverified.

Managed OpenClaw: flexibility comes with more ownership

Managed OpenClaw moves a different responsibility to the provider: operating the runtime. We think it earns its place when chosen models, tools and extensibility matter enough to justify configuration.

Agent 37's personal plans illustrate what that purchase does and doesn't settle. The free instance sleeps after 15 idle minutes; paid plans describe persistence. Model access can be supplied separately or bundled with limits. Downloadable workspace files offer an exit route, without proving that active logins and schedules migrate intact.

We wouldn't equate those hosting features with a finished assistant. The host can keep a system online while the owner still defines its useful job. OpenClaw's execution-approval documentation also doesn't establish a particular host's deployed defaults.

Our assessment is favourable where flexibility is the purpose. It is less compelling as an explanation of how an ordinary recurring errand gets finished. That distinction is about the work being purchased, rather than a judgement on performance we haven't tested.

Our opinion: the advantage comes from what the agent can reach

The clearest pattern in our investigation was the relationship between access and useful responsibility. Grok can reach the conversations on X. Dots can coordinate work in Codex and ChatGPT Work. Instinct makes purchases and personal errands part of the offer. Muse reaches business systems. Those connections give each proposition substance.

We don't think the category will be understood well through a single intelligence ranking. An agent can be capable yet poorly placed to finish a particular job. Our stronger expectation is that people will develop preferences around the work an agent reliably carries, the context it retains and the relationships it can act within.

That has consequences beyond the assistant companies. If customers arrive with an agent to buy, book or resolve something, the receiving business has to finish its side of the transaction. Our PAP and PACT investigation examines that emerging interface. The interesting commercial question is how far the customer's chosen assistant can actually take the job.

For now, we see distinct reasons to take several of these products seriously. Our next evidence should be equally specific: a sourced Grok investigation, a dot coordinating a real implementation, or an Instinct order carried through an exception. Those observations could strengthen or change our opinion. The documents gave us a reason to ask those questions; they haven't supplied the results.

Primary sources checked 9 October 2026. No comparative benchmark, purchase, account-connected trial or vendor interview was conducted. Product capabilities are attributed; assessments and illustrative assignments are ours. About Agentic Economy.