For a research job, is a browser with no logins enough, or do you need your real session?

For a research job on public pages, a browser with no logins is enough. You do not need your real session, and you should not give it to the agent. Your real session is your accounts, your cookies, and whatever passwords the profile can fill. Use it only when the fact you need is behind a login, and even then do not hand the session over. Open the page yourself and paste the part the research actually needs. A logged-in agent is not more thorough. It is you, clicking, on every site that profile can reach.

If the public page is enough for a stranger, it is enough for the agent. If it is not, the missing piece is a paragraph you can copy, not your whole browser life.

What my daily profile added that I did not ask for

I sent an agent to a vendor’s docs to answer how their staging hostname is chosen. The docs are public. I pointed it at my everyday browser because that window was already open and I did not want to set up a profile. The agent followed a “see your plan” link in the nav, which the docs show to everyone and which, for me, opened the billing dashboard. The note it wrote included my plan name and a seat count. Neither was the question. Both were available because the session was real.

The research was worse, not better. The plan name crowded the answer, and I had to check whether the hostname rule it quoted was from the docs or from a banner in the dashboard. A logged-out window on the same docs page produced the hostname rule and nothing about seats. That was the whole job. The real session had contributed a leak and a distraction.

I deleted the note, signed the agent’s browser out of everything, and switched it to a profile I had created with an empty password list. The next research run could not open the dashboard. It said the link wanted a login and stopped. Stopping was the correct result. I already knew my seat count. I did not need an agent to read it in order to explain a public doc.

What “no logins” has to mean

A separate profile, not your daily one with the accounts signed out for the afternoon. Sign-out is reversible by a restore, a sync, or a click on a remembered account. The agent profile should never have signed in. Check the password settings and the list of signed-in sites before the first run. If sync from your personal browser is on, turn it off for that profile. Sync is how the real session comes back under a new icon.

No password manager extension in that profile. No saved cards. No “continue as” buttons you have accepted before. Those are logins that do not look like a settings page. I open the profile myself, visit a site I have an account on, and confirm I see a logged-out page. If I see my name, it is not the research browser.

Cookies from the research itself are fine and should die with the profile or at the end of the run. A cookie that keeps you in a docs consent banner is not your bank. A cookie that is your bank is a login. When I am unsure, I delete the profile after the job. Profiles are cheap. Sessions are not.

When the public page is not the fact

Some questions live behind a login on purpose. Your invoice, your repo’s private README, a ticket in a tracker. A logged-out browser will stop, and that stop is not a reason to attach your real session. It is a reason to decide whether the agent should see that fact at all.

If it should, paste it. Copy the ticket text, or the paragraph of the private doc, into the prompt. You have chosen the bytes. The agent still has no session, so the next link it opens cannot wander into billing. Pasting is slower and it is bounded. The real session is unbounded: every tab that profile can restore, every account that shares the cookie jar.

If the research is “what does this public page say,” and the page walls part of it, quote the public part and say the rest was walled. Do not log in to be complete. Completeness is how my seat count entered a hostname note. A partial answer you can trust beats a full answer that mixed a dashboard into the docs.

What a real session lets a page do

The agent is reading untrusted text with your cookies attached. Any site it opens can be served as you. That includes a docs site that has a surprise redirect, a search result, or a page that used to be fine. I am not going to describe how a page pushes an agent. The outcome is enough: actions and reads that your login makes possible, in a prompt you meant to be research. A logged-out profile can be misled about public facts. It cannot spend your session, read your inbox tab, or fill a saved password. I accept wrong summaries. I do not accept the session as a research tool.

People say they will watch it. Watching does not scale to a run you started and then left, and it does not catch a second tab. The profile is the control that works when you are not looking. If you cannot leave it, you do not have a research setup. You have a supervised login, and you should do the clicks yourself.

The habit that puts the real session back

“Just this once, the docs are behind the customer login.” Then the profile has one login, and the next task reuses the profile because it is the agent’s browser. One login becomes the jar. I do not do the once. I log in as myself, in my own window, copy the paragraph, and leave the agent profile empty. If that is too slow for the volume of private research, the volume is the problem. A dedicated read-only account with nothing else in it is the only middle step I tolerate, and I close it when the project ends. My daily profile is never that account.

I also refuse the agent’s offer to “use the browser you’re already in.” That offer is the real session with a shorter sentence. The browser I am already in is where I read mail. Research does not get it.

Search pages and “open the top result”

Research often starts at a search box, not at a doc you already trust. A logged-out search is still the right browser. The top result can be an ad, a lookalike domain, or a forum where someone is wrong. Your real session does not make the top result truer. It makes the click more expensive, because the lookalike may get a cookie or a filled login if the profile is helpful. I let the agent collect titles and URLs from the logged-out window, and I open the one I meant. If I do not have time to open it, I do not have a research result yet. I have a list. Lists are what a no-login browser is good at. Judgment about which domain is yours stays with you, especially when two domains differ by a character.

Downloads are the same split. A research job that “grabs the PDF” should grab it in the empty profile, and you should open the file only after you look at where it came from. A download inside your real session is a file plus whatever that site just learned about your account. I would rather a PDF in a throwaway downloads folder than a smoother fetch that rode my work login. If the PDF is behind the login, I download it myself and hand the agent the pages I am willing to have summarized. The agent does not need the site. It needs the pages.

The line for a research job

Public question, logged-out profile, empty passwords, no sync. Private fact, you paste the excerpt, the agent still has no session. Your real browser stays yours. If a result depends on being you, it is not a research convenience. It is an account grant, and research is not the reason to make it.

A browser with no logins is enough for the job you can defend. The jobs that seem to need your real session need a paste, or they need you. They do not need an agent wearing your cookies. If you are about to reuse the window where your mail is open, stop and make the empty profile first. The docs will still be there in five minutes. Your session should not be the thing that made the research possible. Thoroughness is a public page plus a paste you chose, not a browser that already knows who you are and can walk into the next account from the same cookie jar.