Comment Pins in the Browser: Hermes Desktop 0.21 Turns Websites into a Sandpit Editorial Desk

Most screenshots die in a folder called Screenshots. That is how I worked for years: look at the page, take a screenshot, paste it into ChatGPT, ask the agent to change something, and watch it guess what I meant. A screenshot shows a picture, not the document. The agent guessed the selector of the button, estimated the class of the header, and I fired correction rounds like free tickets to a bad concert. Hermes Agent has repaired exactly that spot in version 0.21. Not with a new model, not with more context, but with an honest observation: if an agent does not know the page, it should be able to ask the page instead of guessing from an image. The desktop client now ships a comment mode in its browser. I click Annotate, click an element or draw a box around an area, and type my note. Every saved comment stays on the page as a numbered pin.

The part that separates this from a screenshot pen: a pin carries the CSS selector, the markup of the element and the computed styles that matter for layout. The agent receives the source, not the photo of my page. It finds my element in the code instead of guessing it from pixels. Passwords and fields that look like keys are redacted before the page leaves the client. That is the kind of detail you only miss once you have had it. I run this version on the Pixel cyberdeck, a used Google Pixel 6a where Hermes runs as a gateway with cron and the Telegram bridge, while the desktop client runs on the MacBook next to it and the work travels over Telegram between the machines. The test was set on September 11: give feedback to my own agent family with comment pins, without touching a line of code myself, and check the version against the official release. What follows are the numbers, the limits and the moments where the pin workflow was faster for the first time than everything I used before.

The release notes for 0.21.1 (tag v2026.9.7, released September 7) carry the core in one sentence: saving never sends a turn. The pins stay on the page, collected as a numbered list. Only when I say Add N comments do the pins become a handful of work packages: each pin arrives with a cropped screenshot and a short prompt naming the comment. Twenty points of criticism turn into a few batches instead of twenty separate tasks. Why that matters shows up in the bill. A turn is a request on my setup, and requests are rationed: my cloud access runs over Ollama in a five-hour window with a weekly cap. Twenty pins as twenty separate tasks would have sent the full context twenty times. The batching logic groups pins by the region of the page they sit in, and because the groups are separate DOM subtrees they usually touch separate files, which is exactly what makes it safe to hand them to parallel workers.

The numbers for the release window come from the official release notes on GitHub and were counted there by a third party: a local clone yielded 5,140 non-merge commits, exactly one more than the snapshot number, which fits the release commit delay. I recount such things because I once built a blog post on a wrong word count and never want to relive that. The gap of one commit is not a discrepancy, it is the proof that the number comes from a real git window.

My old loop for website changes looked like this: take a screenshot, paste it into the app, write text around it, wait for follow-up questions, answer the follow-ups with more screenshots. Four to six messages for a change that was clear in my head within ten seconds. With pins it is one action: I mark the element, type the note, and the agent already holds the selector and the styles. The follow-up question of whether I mean padding or margin disappears, because the computed styles contain the answer. The second difference sits in the desktop client itself. Version 0.21 also shipped structured session controls: sessions can be created from the new session dialog by drag and drop and sorted into profile groups, and there is an import view for transcripts of foreign coding agents. That fits what I have been building all summer: a small workshop where several agents sit on separate tasks in parallel, instead of watching one generalist stall in sequence. I orchestrate, the pins deliver the raw material, and the session controls keep the workshop sorted.

Honesty applies: my test was an everyday test, not a lab study. I counted how many messages the loop needs, I checked whether pins survive a reload, and I watched whether the agent hits the intended elements. It did so reliably in my runs, because the pins carry the structure. What I did not do: systematic benchmark series over a hundred pins or a blind comparison against other tools. I hand in the methodology openly instead, so anyone can size my test verdict. The release trail is verifiable: tag v2026.9.7 for 0.21.1 and v2026.9.11 for 0.21.2, both in the official release channel.

The industry is debating which interface agents deserve. Rabbit R1, Humane AI Pin and Omi tried their own hardware in 2024, Plaud NotePin tried again, and all of them failed at the point where an extra device annoyes the daily routine instead of helping it. The more interesting question is not which new hardware deserves agents but which tools people already carry open. A phone is such a tool, and my Pixel 6a now carries a complete agent operation. The comment mode takes the same path from the other side: it does not bet on new hardware, it builds on the surface web work already has, the browser. I do not have to learn to talk to an agent. I only have to do what every design round already does: point at the spot that bothers me and say what is wrong. The rest is mechanics, and the agent already holds that in the selector and the styles. A sentence stuck with me while preparing this post: a screenshot shows how the page looks. A pin shows what it is. Whoever does web design knows that the distance between those two statements is the whole job. When an agent does not have to bridge that distance, it does not save seconds, it saves rounds. And rounds are the expensive part, because every round is a turn.

First limit: the comment mode sits in the desktop client, not the phone. I steer remotely over Telegram, and the pins are created at the desk while the MacBook is open. For my workflow that is no problem, because web design happens on the big screen anyway. It is not a mobile solution, and whoever expects the whole loop from the phone is reading it wrong. Second limit: pins do not replace the proper brief. When I say the button should look nicer, the agent gets a selector but no taste. The value of the feature sits in the precise spot, not the precise intent. I kept my notes deliberately short in the test and still made the design decisions myself. The agent builds what I pin, not what I mean. Third limit: the comment mode did not replace my sense for the red lines, it sharpened it. I now see layout inconsistencies faster because the pins are numbered and sit side by side. That is good for the work, but it also means I formulate more corrections than I would have typed before. Whoever thought agents make feedback obsolete confused the part that stays: the judgment.

This blog page here is the next test candidate. I plan to run a revision of the comparison series entirely through pins: every heading, every call to action, every table gets a pin with a short note, and I let the changes run as a batch. Then I will put that loop next to my old screenshot loop and compare the message counts. If the number drops and the quality holds, I will write the update as its own post, with concrete numbers instead of feelings. In parallel I am looking at how the import view for foreign transcripts plays with my workshop. I have agents on different devices, and their logs land in different formats. If the desktop client can read them, my agent family becomes an agent family with a shared photo album. That is a test for the coming weeks, not a promise. What I can say today is only what I measured: version 0.21 moved my loop from the screenshot to the element, and that was the right step.