Key findings on the state of personal AI agents
The October capabilities register covers 19 third-party plan entries across 17 products. ChatGPT Work Plus/Pro and Claude Pro/Max each count as one product, with a capability counted if any listed paid plan documents it. Gemini scheduled actions and Gemini Spark remain separate products because Google documents separate task systems.
- 2 of 17 products document memory export. View matching register rows.
- 3 of 17 products document payment approval or a handoff to the user. View matching register rows.
- 10 of 17 products document send approval settings. View matching register rows.
- 8 of 17 products document user model choice. View matching register rows.
Approval counts cover documented policies, including adjustable rules, skipped confirmations and payment handoffs. They measure the scope of the documentation, not mandatory approval for every action.
Our own product, for transparency: OperatorNest’s row lists our published claims for approvals, memory export and model choice. We exclude it from every finding above.
Memory export is documented in 2 of 17 products
Checked October 3, Meta’s safety page says Muse’s memory files can be inspected, edited and downloaded, while Claude’s export guide explains how to copy saved memory into a local file for backup or migration. Both offer an exit.
The other 15 products’ cited pages leave memory export open, a question to settle before an agent accumulates months of preferences and decisions. Ask for the file. A vendor may offer a feature elsewhere in its documentation, so use the matching register rows as a starting point for the demonstration.
Export also needs an import test. In an illustrative supplier watch, the agent remembers five source pages, a stock threshold and the instruction to hold every purchase. Moving those details to a new service should preserve their meaning. A readable file helps you leave, but the next agent still needs to follow it.
Payment controls appear in 3 of 17 products
Muse, Grok Bot and Pally document payment approval or a handoff to the user. Before you connect a payment method, compare the three workflows in the payment rows, where each policy links to the vendor’s explanation of who approves the purchase, what the agent can execute and when we checked the source.
Muse describes a purchase approval card and adjustable caution. Grok Bot’s security page, checked October 3, says the Bot hands payment steps to the user. Pally’s action policy, checked October 3, describes a single-use card capped at the approved amount. An approval card, a user takeover and a capped payment credential put the boundary in different places.
I would ask the vendor to show the full sequence: prepare the order, change its quantity, reject it, then inspect the order history. A checkbox labeled “approval” tells you little about which amount and seller the permission covers. Change the proposal.
Send policies appear in 10 of 17 products
Ten products describe send approval settings, including policies that allow sending without another confirmation. In the pages we checked October 3, Lindy’s policy says external sends and writes wait for approval, while Manus’s schedule controls let trusted workflows skip confirmations before sending, publishing or posting. Both count.
The send rows include conditional rules in Zo Computer and risk-based review in Grok Bot. A buyer should identify what grants authority for tomorrow’s run. Permission for one draft is a different decision from permission for a tool or every future occurrence of a task.
We left Perplexity Computer out of this count: its cited permission page covers SharePoint and OneDrive writes. Gemini Spark’s cited help page describes confirmation before web browsing. Neither source specifies a send approval policy. This narrower count keeps a general permission prompt from becoming evidence of a held message.
User model choice appears in 8 of 17 products
Eight products document user model selection in their model-choice rows. We count a picker within one provider as well as controls that connect another provider, because the question is whether the user selects the model for their work rather than whether the vendor lists several models it uses internally.
Manus Flex, checked October 3, documents a connected provider, model and reasoning setting with your own key. Claude’s scheduled-task guide offers an optional model choice during setup. Those choices have different reach. Switch, then rerun the job. Inspect whether its instructions and saved context remain available.
Grok Bot illustrates the value of checking the source. Its FAQ really does name paid Cursor plans and linked SuperGrok subscriptions, and its security page says model selection is managed. We kept the unusual plan cell after checking xAI’s own pages on October 3. Subscription eligibility gives no evidence of a user model picker.
Five checks buyers should demand
Use one bounded task to test the controls together. Example assignment: every weekday at 6:30, check five supplier pages, draft a stock update, and prepare a reorder if stock drops below ten units. Deliver the brief by 7:00, hold every external message and purchase, and identify every page that could not be read. This is an illustrative test.
- Hold a send. Give the agent a draft addressed to your own test inbox. Leave the approval request unanswered. Check that no message arrives, that you can edit the recipient and body, and that a later recurring run still holds its send. Record the permission setting used.
- Hold a purchase. Ask it to prepare a reorder with a fixed quantity and maximum total. Inspect the seller, taxes, delivery and amount before approving. Change the quantity after the proposal and require a new approval. A rejected purchase must leave no order behind.
- Export the memory. Add a harmless test preference, correct it, then export the stored context. Find the corrected value in a file you can read outside the product. Check whether account context, task instructions and preferences have separate exports. A chat transcript does not settle this check.
- Choose the model. Select a different supported model for the same task and verify which choice ran. Check that the instructions and memory remain available. If selection is managed, record that as a buying constraint.
- Inspect the receipt. Make one supplier page unavailable during the example task. The brief should report four successful checks and one failure, with source links and times. Its record should distinguish prepared drafts, approved actions and completed actions. A polished answer that conceals the failed page fails this check.
How we built this
We checked the vendor pages cited in the October JSON and CSV, then counted official documentation across 17 third-party products, with sources and checked dates attached to every cell. These are documentation counts, not results from running the same job on every product. Run pnpm exec tsx scripts/agent-findings.mjs to reproduce the findings, product groups and source-file hash.