Criteria for choosing the best personal AI agents
I think a booking demo is a poor way to choose a recurring agent. A weekday inbox job exposes the parts that matter: a missed connection, an exhausted allowance and a message that must wait for you. I would choose around that job before moving the rest of my week.
I compared vendor documentation on 2 October 2026, without running the same task on every product. We make OperatorNest.
- Works while you’re away: a schedule, an event trigger and a local desktop session have different requirements. Ask whether the task runs with your devices off.
- Channels: distinguish a place you can message the agent from an account it can read. Reading Slack and taking requests there are separate capabilities.
- Model choice: separate a user-facing picker, automatic routing and bring-your-own keys or subscriptions. Ask whether switching preserves history.
- Approvals: identify the exact send, purchase, deletion or publication that pauses. Check whether a recurring task inherits an earlier grant.
- Memory and export: ask what is stored, what you can inspect or correct, and what you can take to another service. Deleting an account and exporting memory solve different problems.
- Receipts and logs: look for sources checked, actions taken, approvals and failures. An enterprise audit log may have different access requirements from your task history.
- Availability and price: check your country, account type, device and required plan. Monthly USD figures below are published prices, with usage charges or separate subscriptions called out where documented.
For dated capability cells and retained monthly snapshots, read the paid AI agent capabilities register. Its downloads keep the source and checked date attached to each detail.
The October state of personal AI agents report counts what those 20 entries establish and gives five checks for sends, purchases, memory, model choice and receipts.
A shortlist by job
Table legend: A dash means the cited page leaves that detail open. Each row names at most one question to resolve for its suggested job. Facts and prices checked 2 October 2026. Sources and eligibility follow below.
| Product and fit | Background work | Where you reach it | Model control | Approvals | Memory and record | Published price |
|---|---|---|---|---|---|---|
| Meta Muse: WhatsApp admin | Schedules and events | Apps, web, WhatsApp | Meta model; BYO: — | Cards; adjustable caution | Readable, editable and downloadable memory; activity log | Free for most needs; paid amounts: — |
| Poke: text delegation | Recipes; paid automations | Messages, WhatsApp, Telegram | — | Ask: does every send wait for my OK? | Memory; export and action-log coverage: — | Free; Pro $19/mo; Ultra $199/mo plus usage |
| Instinct: contextual follow-ups | Proactive follow-ups | Text and calls | — | Ask: can I require approval before sends? | Context collection; memory tools and receipts: — | — |
| ChatGPT agent / Work: project tasks | Schedules and events | Desktop, web, mobile | OpenAI picker; external BYO: — | Check-ins and app permissions | Memory controls; enterprise logs | Plan dependent; Pro $100/$200/$500 monthly |
| Gemini Spark: Google tasks | Cloud schedules; local needs device | Gemini mobile, Mac, web | — | Web confirmation. Ask: which scheduled actions need my OK? | Connected context; memory export and receipts: — | US Pro $19.99/mo; Ultra $99.99/mo |
| Grok Bot: persistent computer | Routines and events | Desktop, mobile; team Slack | Managed selection; no picker | Configurable review | Shared files; enterprise logs. Ask: can I export memory? | Bundled paid access; allowances vary |
| Lindy: inbox and meetings | Scheduled briefs | Slack, iMessage, connected inbox | Per-task choice; BYO: — | External sends/writes wait | Workspace context; action logs. Ask: can I export memory? | From $29.99/user/mo |
| Manus: BYO-key projects | Schedules and events | Web, Telegram | Flex BYO keys and model controls | Send/publish confirmations can be skipped | Persistent context/files; receipts: —. Ask: can I export saved preferences? | Free; Pro from $20/mo plus Flex inference |
| Genspark: cross-channel Claw | Cloud schedules and monitoring | Workspace; Claw chat channels | Claw picker; BYO: — | Claw external-action permission | Cross-channel memory; no chat-history export. Ask: does each run have a receipt? | Plus $24.99/mo; cloud Claw costs extra |
| Perplexity Computer: research | Recurring and conditional tasks | Web, desktop, mobile, email | Multi-model; BYO inference: — | Check connector grants | Brain editing on eligible plans; receipts: —. Ask: can I export memory? | Pro $20/mo; Max $200/mo |
| OperatorNest: independent recurring work | 24/7 schedules and watches | Chat apps, email, web | Any provider; own keys/subscriptions | Consequential actions wait by default | Read/edit/export/delete memory; action receipts | Get access |
Best for delegation from messages
Meta Muse
Meta’s product guide describes scheduled and event-triggered work, readable and editable memory files, an activity log and approval cards. Its business announcement adds connected business accounts and says sends, publishing and spending need approval.
Limits: Meta doesn’t describe BYO models. Its security guide says you can download memory files; importing them into another agent is a separate step. Caution settings change approval behavior. Availability is US and Canada. Price, checked 2 October 2026: free for most needs, with paid subscriptions whose amounts Meta doesn’t state. Our Muse comparison.
Poke
Poke handles more than messages: its FAQ describes Recipes, email/calendar connections, custom tools and website creation, while the published Pro plan adds background automations for work you want to repeat without starting a new conversation each time.
Limits: ask Poke about model choice, memory export, the AI’s approval gate, country eligibility and action logs. Its FAQ describes controls for human help separately. Price, checked 2 October 2026: Free, Pro $19/month, Ultra $199/month with usage charges. Our Poke comparison.
Instinct
Instinct’s site describes proactive assistance through text and calls, with a direct text entry point and optional connected accounts and device information for contextual follow-ups and bookings, which makes it a candidate for personal tasks where a call may help finish an errand.
Limits: its terms authorize connected-service actions without a universal send confirmation. Disconnecting leaves indexed copies; request removal separately. Ask about memory inspection/export and task receipts. Availability and price, checked 2 October 2026: age 18+; model choice, public prices and eligible countries are unspecified. Our Instinct comparison.
If these messaging assistants are your shortlist, Poke vs Instinct compares eligibility, recurring work, approvals and export with sources checked October 3, 2026.
Best for work inside an existing platform
ChatGPT agent
OpenAI’s agent help directs users from the original agent mode to Work, where scheduled and event-triggered project tasks use an OpenAI model picker across eligible desktop, web and mobile surfaces while keeping the job inside ChatGPT. The Work admin FAQ explains app permissions and enterprise observability.
Limits: access depends on plan and workspace. OpenAI doesn’t describe external-model BYO or cross-provider memory portability; memory controls cover review and deletion. Price, checked 2 October 2026: Pro tiers are $100/$200/$500 per month. Work’s minimum plan is a separate eligibility check. Our ChatGPT comparison.
Gemini Spark
Google’s Spark help describes schedules, connected Google context and web-action confirmation, with cloud work continuing while devices are off and local sessions requiring an awake computer, so check which environment will run the job before closing the lid.
Limits: Google’s help requires a personal account, age 18+, eligible Pro/Ultra and Keep Activity, excludes EEA, Nigeria, Switzerland and UK, and leaves external-model BYO, memory export and full approval coverage open. Its pages disagree on business eligibility. US price, checked 2 October 2026: Pro $19.99/month; Ultra $99.99/month. Our Spark comparison.
Grok Bot
The FAQ documents desktop/mobile access, routines and shared persistent files. Team Bots adds Slack access on Teams/Enterprise. Model selection is managed; there is no customer-facing picker.
Limits: bots share the account’s computer, files and logins. Roles aren’t security boundaries. Security documentation describes configurable review, enterprise logs and opt-in action recording; standalone memory export is unspecified. Price and access, checked 2 October 2026: bundled with paid individual Cursor plans and Teams, with eligible SuperGrok access; allowances vary. Ask about country eligibility. Our Grok Bot comparison.
Best for connected work and research
Lindy
Lindy’s pricing FAQ describes Slack/iMessage access, per-task model choice and scheduled reports for inbox and meeting work, with read-only work on approved sources proceeding automatically while external sends and writes wait for approval, a boundary worth checking against your follow-up job. Its security page says each action is logged.
Limits: credits can pause work. Self-serve access is advertised; ask Lindy about countries, BYO inference and memory editing/export. Price, checked 2 October 2026: Plus $29.99/user/month, Pro $99.99, Max $199.99, with plan-specific credits. Our Lindy comparison.
Manus
Manus Flex lets you choose provider, model and reasoning settings. Manus 2.0 adds event-triggered Automations and a persistent Cloud Computer; Telegram delegation is available across tiers.
Limits: scheduled-task settings can skip send/publish confirmations. Ask about general memory export, receipt coverage and country eligibility. Price, checked 2 October 2026: Free includes two scheduled automations; Pro starts at $20/month. Flex inference is billed by your provider; other resources consume Manus credits. Our Manus comparison.
Genspark
Genspark Claw help documents model selection without clearing history, cross-channel memory, WhatsApp/Slack/Teams/Telegram access, schedules and Heartbeat monitoring. It says external sends and posts require permission.
Limits: local Claw needs the computer/app running; cloud Claw needs a separate subscription with an unspecified price. Chat-history export is unavailable. Ask about BYO inference, receipt coverage and country eligibility. Price, checked 2 October 2026: Free; Plus $24.99/month; Pro $249.99/month, before the cloud-computer charge. Our Genspark comparison.
Perplexity Computer
Computer help describes multi-model research, background jobs and conditional triggers across web, desktop and mobile. Brain gives eligible Max/Enterprise Max users source-linked memory they can inspect, edit and delete.
Limits: ask Perplexity about memory export, BYO inference and receipt coverage. Review connector grants before scheduling work, including which consequential actions pause. Price, checked 2 October 2026: Pro $20/month; Max $200/month. Paid-plan access is documented; ask about countries. Our Computer comparison.
Best for independent recurring work
OperatorNest
OperatorNest runs schedules and watches from chat apps, email or the web. You can choose a model from any provider and bring your own keys or subscriptions. Memory can be read, edited, exported or deleted; switching models keeps it. Consequential actions wait for approval by default, and receipts record actions and decisions. Specific tasks can have scoped pre-approvals.
Limits: receipts need review; research can still be wrong. Keep a job in a product you already pay for when it covers your sources and approval rules. OperatorNest needs its own account connections and task setup; moving a job means recreating those instructions.
Example: one job to check before handing over more
For example, Sam wants a weekday inbox follow-up brief at 7:00: “Read the connected inbox, list replies I owe, draft three nudges, and hold every send for my approval. Show which messages you checked and any failures.”
Close the laptop for one run, reject a draft, inspect the record, then correct a recipient preference and ask to export that correction. Follow the whole job. If a missing inbox connection produces an empty brief without reporting the failure, fix the task before making it recurring. That check answers more about this job than a polished booking demo.