Compare four personal AI agents.

Start with eligibility rather than capability. A product can sound suitable but still be unavailable to your account, region or organization. A waitlist invitation and a published feature announcement are different things. For self-managed software, installation is only the beginning: model and integration access must also be arranged.

Where work runs determines what stays available when you close a laptop, which files the agent can reach and who maintains the environment. A hosted computer can continue independently, while a connected local machine must meet its product's availability requirements. A local client does not establish that model processing stays local.

Included access is not unlimited work. Background checks, model calls and longer assignments can consume allowances, and sharing a pool with another product changes the practical budget. Self-managed software removes one kind of product fee but can add model, hosting and maintenance costs. Compare the cost of a completed task, not just a subscription headline.

A scheduled task, a periodic check and an event-triggered routine wake up in different ways. That affects how quickly something is noticed and what can happen while you are away. Define the expected output and review boundary, and do not use a periodic personal assistant as an urgent-response guarantee.

Browser tools matter when a task requires interacting with a website rather than reading an answer. Being able to watch or take over that browser can help resolve a blocked login or inspect an action. Tool availability still does not guarantee that a particular website permits automation or that every action will succeed.

Memory can reduce repeated explanation, but retained context also needs to be managed. Disconnecting an app is not necessarily the same as deleting information already learned from it. Separate conversation memory from files and logged-in browser sessions, especially where several named agents share one computer.

Review the action boundary, not just the presence of an approval button. Ask which operations can happen without you, where credentials are handled and whether a failed or unanswered approval stops the work. In a self-managed system, an allowlist is one control; it does not substitute for isolation of the operating environment.

A missing connector or required interaction can decide the choice before model quality matters. Write down the accounts, files and tools your workflow actually needs, then compare them with the documented limits. Treat an unverified capability as a question to investigate, not as either a promised feature or a confirmed absence.