The Insider Cat

In early summer, I was trying to use OpenHands new UI, dubbed agent-canvas, to test it, to play with it, to work with this incarnation of OpenHands.

It was strangely difficult, though, and I had trouble understanding why. The UI/UX is pretty close to the old OpenHands Web UI, and that could, and had been, okay for me, even if I preferred others over years. Instead, I found myself defaulting to WhatsApp, to my personal agent, smolpaws: my workflow went like “smolpaws get us a sibling in canvas on gpt, then after a turn, switch its LLM profile to deepseek flash, does it work normally?”, then “did you check if the LLM really changed?”. Then maybe “give me a tailscale link to canvas, I want to see this”.

It was smolpaws who was using canvas, he was using the UI itself even, a UI we meant for (human) users. He was not using only the agent-server API, which, of course, is for agents.

One day it hit me: I couldn’t keep track, once I had more than several conversations in flight. smolpaws was keeping track, since it has memory and it has beads for issue tracking.

The Insider Cat

The Insider Cat is an early prototype exploring the experience of using an agent inside canvas, whose job is to know about the other conversations, what the user is working on, what the user prefers, and what they forgot about.

Voice in Real Time

The Insider Cat can speak. I don’t mean transcribing audio and/or reading out loud, I mean using a real-time voice model with full duplex. The voice model delegates the real work to an OpenHands agent via the OpenHands OpenAI-compatible API. It supports both completions and responses, and it forwards it to a regular conversation with an agent.