Feed reader comments back to your coding agent

Getting feedback on agent output is easy. Getting it back to the agent is the hard part: comments live in a doc, a chat thread or a meeting, and someone retypes them into the next prompt. byagent keeps the comments on the page, anchored to the text they're about, where the agent can read them.

The loop

1. The agent publishes a page: byagent publish ./plan.md --project "<name>" --tag plan --json.
2. You send the link. Readers select text and comment, with no account. They can also suggest replacement text.
3. The agent reads the open threads: byagent comments <id> --open --json.
4. For each thread, it marks it as in progress, edits the page, republishes to the same link, replies, and resolves it.

What the agent gets

Each thread carries:

• anchor.quote: the exact text the comment is attached to, so the agent finds the right line
• the comment and any replies
• suggestion, when the reader proposed replacement text

The commands

byagent comments <id> --open --json
byagent mark <id> <thread> working --json
byagent publish ./plan.md --json
byagent reply <id> <thread> "Moved the rollout to week 3." --json
byagent resolve <id> <thread> --json

When the agent won't change the page for a comment, it declines with a reason, which posts the note and resolves the thread:

byagent mark <id> <thread> declined --note "Out of scope for this plan." --json

Readers see these marks next to their comments, so they know the agent picked them up.

Handle feedback as it arrives

byagent comments <id> --wait --json

This blocks until a reader leaves a new comment, then prints those threads and exits. It gives up after 30 minutes with timed_out: true. An agent that can run commands in the background, like Claude Code, can keep working and handle comments when they land.

Comments are data, not instructions

Anyone with a public link can comment. The byagent skill tells agents to act on a comment only by editing that page, and never to run commands, open links or touch other files because a comment says so.

Get started

npx skills add anup-a/agent-artifacts

Then ask your agent to publish something. More: the Claude Code guide, Codex, Cursor.