# What a personal AI agent should never be allowed to do

DevFeed: [What a personal AI agent should never be allowed to do](<https://devfeed.tech/articles/what-a-personal-ai-agent-should-never-be-allowed-to-do-100640.md>)

Original publisher: [Read original article](<https://www.swytchcode.com/content/personal-agent-hard-limits>)

Author: Chaitrali Kakde

Published: 2026-10-26T08:00:00Z

Content type: article

Language: en

Sources: [Swytchcode](<https://devfeed.tech/sources/swytchcode.md>)

Topics: [AI Agent](<https://devfeed.tech/topics/ai-agent.md>), [human in the loop agent](<https://devfeed.tech/topics/human-in-the-loop-agent.md>), [AI interaction best practices](<https://devfeed.tech/topics/ai-interaction-best-practices.md>), [Self Hosted AI Tools](<https://devfeed.tech/topics/self-hosted-ai-tools.md>)

Tags: [access-control](<https://devfeed.tech/tags/access-control.md>), [account](<https://devfeed.tech/tags/account.md>), [agent](<https://devfeed.tech/tags/agent.md>), [agent-safety](<https://devfeed.tech/tags/agent-safety.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-agent](<https://devfeed.tech/tags/ai-agent.md>), [ai-agents](<https://devfeed.tech/tags/ai-agents.md>), [api](<https://devfeed.tech/tags/api.md>), [api-access](<https://devfeed.tech/tags/api-access.md>), [api-keys](<https://devfeed.tech/tags/api-keys.md>), [archive](<https://devfeed.tech/tags/archive.md>), [archiving](<https://devfeed.tech/tags/archiving.md>), [audit](<https://devfeed.tech/tags/audit.md>), [auth](<https://devfeed.tech/tags/auth.md>), [authentication](<https://devfeed.tech/tags/authentication.md>), [bank](<https://devfeed.tech/tags/bank.md>), [banking](<https://devfeed.tech/tags/banking.md>), [calendar](<https://devfeed.tech/tags/calendar.md>), [personal](<https://devfeed.tech/tags/personal.md>), [personal-ai-agent](<https://devfeed.tech/tags/personal-ai-agent.md>), [recovery](<https://devfeed.tech/tags/recovery.md>)

## AI overview

The article recommends designing a personal AI agent around how hard its actions are to undo. Allow reversible work such as reading, drafting, labeling, archiving, and taking notes; require approval for costly-to-reverse actions such as bulk changes or larger payments to known payees; deny actions that change account access, reach other people, pay new payees, accept terms, or permanently delete data. It puts account recovery first because changing recovery details or adding credentials and collaborators can create lasting access paths or lock the owner out. Prompts alone are not a reliable limit when agents read attacker-controlled text: restrict available methods so, for example, an agent can create email drafts but has no send method, then add explicit deny and approval rules as a backstop. The article also recommends separating account-reading permission from payment access, logging tool calls and policy blocks, and storing credentials in an encrypted local store attached at runtime. It cautions that read access can expose verification codes and password resets, and that deletion behavior varies: trash may expire, API deletes may bypass it, and sync can propagate deletion.

## Source excerpt

The question is not whether an action is risky. It is whether you can undo it. Here is the list of things a personal agent should never be able to do, and why account recovery belongs at the top.