# Share What Antigravity Produced — Walkthroughs That Leave the Editor

Canonical: https://commareports.com/agents/share-antigravity-output
Published: 2026-09-09

> Antigravity writes implementation plans and task walkthroughs into its Artifacts pane, where only you can read them. Publish the walkthrough to a URL your reviewer can open and comment on.

# Share what Antigravity produced

Antigravity's distinguishing habit is that it shows its work. The task plan,
the implementation walkthrough, the browser recording of the thing it just
verified — they land in an Artifacts pane as a document you are meant to read
rather than a diff you are meant to skim.

Which makes it strange that the document is the one part of the workflow with
no address. The code goes into a branch anyone can pull. The walkthrough that
explains the branch stays in a pane on your machine, and gets relayed to the
reviewer as a paraphrase.

## Let the agent publish it

Attach Comma's [MCP server](/mcp) with a [scoped token](/docs/api-tokens), then
put the standing instruction in the repository's `AGENTS.md` so it applies to
every task in the workspace without being restated:

```md
## Publishing walkthroughs

When you finish a task walkthrough, an implementation plan, or a verification
report, publish it to Comma as an HTML report and reply with the URL.

Keep the report id in `.comma-report-id` and update that report on later passes
of the same task, so the link in the pull request keeps resolving and each pass
lands as a revision.
```

The pull request then links to a walkthrough anyone can open, instead of
carrying a description that says "see the plan in my editor".

## Why a report and not a pane

- **Reviewers have no session.** The whole point of a walkthrough is that
  somebody else reads it. A URL is the only format that reaches a reviewer, a
  PM and an on-call engineer with the same effort.
- **It renders.** Screenshots, inline SVG, Mermaid diagrams and tables all
  survive. The markup is stored verbatim and served inside a sandboxed iframe —
  `allow-scripts`, no `allow-same-origin` — so an interactive walkthrough stays
  interactive without reaching anyone's session.
- **Comments land on the step.** "Step 6 skips the migration lock" stays
  attached to step 6 — see [commenting on HTML](/comment-on-html).
- **The agent reads the replies** over the same MCP server, so review feedback
  becomes the next prompt rather than a relay job for you — see
  [letting an agent respond to comments](/agents/let-an-agent-respond-to-comments).
- **Long tasks get a history.** Each pass is a revision at one address, so
  somebody joining on day three can see what changed on day two.

## Worth knowing

- **Scope the token to `reports:write`** unless you want the agent replying in
  threads, in which case add `comments:write` and nothing else. See
  [scoped tokens for AI agents](/agents/scoped-tokens-for-ai-agents).
- **Video and recordings.** A browser recording is not going inside a 5 MB HTML
  body. Publish the walkthrough and attach or link the recording — see
  [interactive HTML reports](/interactive-html-reports) for what does travel
  well inline.
- **Publish private first.** An agent that just browsed your staging
  environment may have captured more than you want public. Widen the access
  setting after you have read it — see [sharing & access control](/docs/sharing).
- **Nothing re-executes.** A published report is a snapshot. Scheduled
  refreshes are [routines](/docs/routines).

## Try it

Comma is free — unlimited reports, unlimited commenters, unlimited revision
history.

**[Attach the MCP server →](/mcp)**

### Related

- [Where should my agent post?](/agents/where-should-my-agent-post)
- [Share what Cursor produced](/agents/share-cursor-output) ·
  [Share what Kiro produced](/agents/share-kiro-output)
- [Give an agent feedback](/agents/give-an-agent-feedback) · [The API](/docs/api)