Agents talk a lot. Ask for one small change and the answer often starts by reading your own words back to you. With reactions, an agent in OpenHands can answer “wait, actually” with a thumbs up, and simply go do the thing.
The film
A real run in Agent Canvas, recorded with the verify-openhands skill against a live Agent Server, in the Paper theme, with DeepSeek Flash as the agent's model. Drawn with Manim; narrated by Kokoro-82M, an open-weight speech model; the score and the reaction sounds are generated with numpy. Download (MP4, 1:27, 4.5 MB) · Scene script, narration, recordings and music
The idea
Fewer wordsA wall of text is tiring to read. When the person changes course mid-task, a 👍 says “understood” without restating the request, and the agent keeps working.
Real toolsThe agent gets react and unreact tools. A reaction is saved as an event in the conversation, so it stays after a reload or a server restart.
Both waysYou can react to the agent's messages too: a ❤️ on an answer you love. Your reactions are kept the same way, and they never start a run.
What happens in the run
Turn it on. Press Ctrl+K, open Experimental, and switch on Conversation reactions. It applies to conversations you start afterwards.
Ask. “Can you write a little script that splits our team dinner bill evenly? It takes the total and a list of names. Put it in split_bill.py and show me an example run.”
Change your mind. While the agent works: “Wait, actually: not everyone had drinks. Can it take what each person ordered, and split only the shared appetizers evenly?” The agent puts a 👍 on that message and carries on.
A short answer. The new options, one example run, and a line about leftover pennies. You react with ❤️.
Restart, reopen. After control-openhands restart, the conversation opens with the 👍 and the ❤️ where they were.
One honest detail
The demo repository has an AGENTS.md with a short note on how the person likes to work. In four earlier runs without it, DeepSeek Flash and DeepSeek Pro handled the change of plan correctly, but none of them reacted on their own: the tool's description suggests 👀 when starting on work and 👍 for a note that needs no reply, and the models did not read a mid-task correction as either. With the note, the agent reacted on the first try and again in the recorded run.
## Working with me
- Keep answers short. I can read the code.
- If I change my mind while you're working, don't restate what I said: react with a thumbs up (+1) on my message and carry on.
Sources
Agent Server: experimental feature switches and the reactions API and tools in software-agent-sdk #6609. Canvas: Settings › Experimental, reaction chips and the picker in OpenHands #19503. Both are open pull requests; the film was recorded from their branches on 9 November 2026 (SDK c4beabb, Canvas verify branch 1e9d355). Everything here is open source, and anyone can rerun it.