An agent typed the right reply, but the Respond button stayed disabled. Better prompting wouldn't fix it.
retained text
I ran into a useful failure while testing browser automation with an AI agent on my Mac. The task was trivial on paper: open a Medium response, type a short reply, and publish it. The agent correctly found the textbox. Its browser tool returned a successful write. The reply was visibly in the editor. But **Respond stayed disabled**. We reloaded the page and tried again. Same result. Then I copied the exact same text and pasted it manually. The button immediately became enabled. That made the distinction obvious: **the DOM looking correct is not the same thing as the application accepting an edit**. Many rich-text editors maintain their own state. Assigning a contenteditable element's text can change what you see without producing the selection/input transaction that the editor expects. A tool can confidently report "value applied" even though the application's actual state is still "nothing to submit." No amount of "be careful," "try another selector," or "verify before clicking" in the system prompt repairs that input path. It has to be addressed in the automation layer and tested against editor behavior. A few lessons from this bug: 1. Treat tool success as evidence about the tool, not automatically about the user's goal. 2. Verify the *next state* that matters: was the submit control enabled, did the save persist, did navigation occur? 3. Test against controlled editors and real sites. Native input fields, contenteditable, and framework-managed rich-text inputs behave differently. 4. When a mutation may already have happened, avoid blindly replaying the final Send/Submit click. An ambiguous outcome needs verification, not optimism. I encountered this while building [Mac MCP](https://github.com/bulutarkan/mac-mcp), my local Mac control layer for agents. We fixed the rich-text typing path and reproduced the original test until the Respond button actually enabled. We did **not** submit a second comment just to prove the fix. I'm curious how other people handle these false-positive successes in browser agents. What do you verify beyond the automation tool's own success response?
document fetched 2026-10-11 00:36 · canonical https://reddit.com/r/PromptEngineering/comments/1x2gauw/an_agent_typed_the_right_reply_but_the_respond