Security
Login walls are part of the system
A password prompt is not a bug in the workflow. It is the point where control returns to a person.
Automation demos make sign-in look like plumbing. Fill the fields, click the button, carry on. Real accounts are messier because the friction is intentional.
A CAPTCHA asks whether a human is present. A passkey asks for possession of a trusted device. A one-time code tests access to another channel. A legal consent screen asks someone to accept terms. An agent should not treat any of these as puzzles to beat.
Prepare everything around the gate
Stopping does not mean wasting the run. Before handing control back, an agent can still:
- open the exact first-party page;
- select the correct non-secret account context;
- prepare the title, description, file, and visibility setting;
- validate the asset and preserve its fingerprint;
- state the one action a person must complete.
The handoff should be narrow. "Please sign in" is weak. "Complete the visible CAPTCHA and leave the upload page open" is useful. It tells the operator what to do without asking them to expose a credential.
Do not solve trust by weakening it
When automation triggers an unsupported-browser warning, repeated retries rarely improve the situation. Neither does disabling browser security, copying session data into random tools, or routing credentials through an interface that was never meant to hold them.
The safer alternatives are boring: use an already authenticated normal browser profile, use the provider's official API with properly stored tokens, or stop for a visible human login.
Resume from durable state
After the gate, the agent should verify the account from first-party state. A form submission is not enough. Look for the account menu, channel page, dashboard, or authenticated API response. Then continue from the prepared artifact instead of rebuilding it.
A good boundary preserves momentum without pretending the boundary disappeared. The agent does everything it can safely do, asks for one unavoidable action, and resumes only after the provider confirms the result.
This note describes a conservative operating rule for account automation. Provider requirements and supported authentication methods vary.